Projects

Project ID 01: 

Project Title:  Explainable AI and Multimodal deep learning for Healthcare Data

Student Members: 

Pranav Nayak, Final Year IT

Ashmit Bharadwaj, Final Year IT

Mentor Name: 

Dr Rahee Walambe Associate Faculty ,SCAAI

Associate Professor, Dept of E&TC , SIT

Objectives and Methodology

In the 21st Century we are moving towards information age, where more and more sedentary jobs are getting created. In this fast paced life mental health is of the utmost importance. Mental health issues can both directly and indirectly affect other parts of the body and capable of hampering your day to day activities and your performance. We therefore aim to help people working sedentary jobs to monitor their working behaviour and stress levels. This will help users understand their own body and then they and take necessary steps to improve their health and performance.

Our methodology here is to develop AI based multimodal fusion techniques for handling multimodal data so that it can be used efficiently for training AI models. Multimodal fusion enables users to interact with computers through various input modalities like speech, gesture, and eye gaze. Though combining different modalities or types of information for improving performance seems intuitively appealing task, but in practice, it is challenging to combine the varying level of noise and conflicts between modalities. Moreover, modalities have different quantitative influence over the prediction output. To generate human interpretable explanations for these AI models using Explainable AI methods. To store these explanations securely in a blockchain so that the user can access these anytime with full anonymity.

Project ID 02: 

Project Title: To Build up a AI based Smart Entry Log System for storing data obtained from live Camera Recording.

Student Members: 

Anusha Gadgil, Final Year BTech (E&TC)

Arjun Thakur, Final Year BTech (E&TC)

Mihir Gohad, Third Year BTech (MECH)

Mentor Name: 

Dr. Rahee Walambe,

Associate Faculty, SCAAI and Associate Professor, Dept of E&TC, SIT

Objectives and Methodology

The ultimate aim of our project is to ease the process of logging details of vehicles entering our campus in real time. We plan to achieve this using Deep Learning Algorithms like CNN and SSD. First, we detect the License plate then recognize it, then we check for 2 wheeler, person and helmet. If it’s a 2 wheeler, all of this data is stored in excel sheet.

Video processing will be done using camera feed to look for vehicles. We have used our XML classifier with OpenCV functions for detecting License plates from the video. Each time the Plate is detected in the video, it is saved as a cropped image to a desirable location where OCR is applied on it. The OCR will recognise licence plate and stored. Once we get new cropped image, the old one gets deleted automatically. We made sure that the plate is detected only once. Therefore, we stored this License plate in text format along with the Timestamp. We have tested around 150 images on OCR and got an accuracy of 90%. Furthermore, we tested License plate and helmet detection algorithms on around 30 videos and got 80 % accuracy and collected 1,000 images for training Helmet detection model & then saved the model. Lastly, we used SSD model for detecting person, two-wheeler with our saved CNN model for detecting presence or absence of Helmet and then stored the data in Excel Sheet.

Project ID 03: 

Project Title: Development of Techniques for Generating Photo Realistic Images from Textual Description

Student Members: 

Sharad Pande

Ritesh Sonavane

Srishti Chouhan

Final Year BTech (EnTC)

Mentor Name: 

Objectives and Methodology

Converting from text-to-image is a fundamental and novel domain of research in computer vision. It could be seen as a reverse task to image caption, aiming to generate photo-realistic images from input textual description sentences. Text-to-image synthesis helps to find the relationship between text and image, traversing the visual interpretation mechanism of the human brain. Besides, it also has numerous application potentials like computer-aided design, art creation, face generation, image searching. Our main goal is to be able to generate faces from its textual description, and we intend to translate text which is in the form of human-written single-sentence descriptions directly into images. We used Stack GAN, PG GAN and AttnGan to achieve this conversion.

StackGAN: This GAN is typically used for synthesizing images from textual description. It breaks down the text-to-image generation process into two stages, Stage-I and Stage-II.

Project ID 04: 

Project Title: Detecting onset of alzheimers using CNN

Student Members: 

Varun Jain (ENTC) - 17070123045

Daryl Jacob Jerrish (CS) - 18070122017

Om Nankar (IT) - 19070124045

Mentor Name: 

Dr. Shilpa Gite

Objectives and Methodology

The objective of this project is to find patterns in MRI images of the patients brains’ to detect onset of alzheimers. Our approach towards the project was using the Convolutional Neural Networks as the dataset contains images. The methodology can be broken down to 4 stages:

Dataset acquisition and processing:
The dataset we took into consideration had to be augmented (data augmentation techniques such as flipping the image, rotating the image) in order to better our findings.
Training the model using CNN:
The data was divided into 4 categories:
Non-demented
Very mild demented
Mild demented
Moderate demented
Usage of feature learning through application of strided convolutions and pooling layers in the model, feature selection and extraction.
Optimization:
Improving accuracy and betterment of results using ADAM optimizer and adding categorical cross entropy for loss computation.
Training and Deployment.

Project ID 05: 

Project Title:- Automated Coronary Artery Disease(CAD) diagnosis using ECG signal and convolution neural network 

Student Members:- Abhinav Mishra second year(CS)

Mentor Names:- Dr Shilpa gite, Hariharan Venugopal 

Objectives and methodology:

The primary objective of the project is to distinguish between cad and non-cad patients using ECG signal of their heart using a custom convolution neural network.

To perform cad classification on ECG signal, first, we denoise using wavelet transform, median filter, and low pass Butterworth filter because raw ecg may contain noise like baseline wander, power line interference, motion artifact etc after denoising ECG segmentation was done to increase the dataset and reduce the effect of the noise and then normalization technique is implemented to normalize the individual ecg segment this is to overcome the problem of amplitude scaling and to remove the offset effect .Custom cnn was used to learn the features from input data and a classifier was used to perform classification.

 

Project ID 06:

Project Title- Emotion Detection from Audio and Video 

Student Members

Chirag Dalvi – Final Year BTech (CS)

Manish Rathod – Final Year (IT)

Arihant Surana – Third Year BTech (CS)

Mentors: Dr.Shilpa Gite & Dr. Shruti Patil

Objectives and Methodology

In times of current pandemic, we have been forced to adopt new digital lifestyles including Online Teaching. Most schools and Universities have shifted their teaching from Offline to Online mode. Though the Study material and the syllabus remain the same but handling an online teaching is still a difficult task as compared to conventional physical classes. This project aims to develop an automated AI tool to create for teachers to understand the emotional state of kids during online school learning to create activities or study methods to bring up their learning moods. To analyse the Emotions from both audio and video, we have divided the project into four categories: DataSet Selection, Visual Facial Emotion Recognition Model and Audio Emotion Recognition Model, Fusion of Predicted data from both Video and Audio Model, respectively. We aim to generate a robust dataset of audio and video clips for implementation. After individual predictions, we will fuse the results and build a software that can detect and classify the emotional state of the end user in real time.

Video

Audio

Project ID 07:

Project Title:  Yoga Pose Estimation

Student Members:

• Jenet R Scaria, Final Year BTech(E&TC)
• Arihant Surana, Third Year Btech(CS)
• Drishti Singh, Third Year BTech(CS)

Mentor:  Dr. Shilpa Gite, Assistant Professor, CS Department, SIT

Objective/Methodology:

With the recent advancements in Artificial Intelligence, computers can now leverage many human-like abilities in their operations using image processing and computer vision technology. One such application is Human Pose Estimation. The poses are distinguished from each other at the joints of body parts. An application of pose estimation which has attracted many researchers is exercise and one of its form with intricate postures is Yoga which is an age-old now famous worldwide because of its many spiritual, physical and mental benefits.

Pose estimation for fitness applications is particularly challenging due to the wide variety of possible poses (e.g., hundreds of yoga asanas), numerous degrees of freedom, and occlusions. The approach used here is the skeleton-based model where a set of joints (keypoints) like ankles, knees, shoulders and limb orientations of a human body are detected from an image or a video frame, which are then utilized to create a skeletal structure of the posture at any instant of time.

The RCNN model has shown the most significant results in Pose estimation problems. The input image goes through the CNN. The feature maps generated are used for the region proposal network (RPN) to generate region proposals, and used for generate feature maps for RoIAlign pooling at later part. The whole network is an end-to-end deep learning network, which is essential for gradient propagation to improve the pose detection accuracy.

Project ID 01: 

Project Title: A Study on the Integration of Explainable AI and Blockchain and its Application in Credit Scoring

Student Members: 

Ashwin Kolhatkar, Final Year BTech (CS) 

Manas Ojha, Final Year BTech (CS) 

Akash Kademani, Final Year BTech (CS) 

Sakshi Kathote, Final Year BTech (CS) 

Mihir Pandya, Final Year BTech (CS)

Objectives and Methodology

Credit scores are an important factor that financial institutions consider when deciding whether to approve a loan or not. The scores are designed to predict the likelihood of repayment of a loan. Many of the current ML/AI algorithms just provide the result on whether a particular customer will default on loan repayment or not. An explanation as to why the loan was denied is usually not available. The major issue with complex machine learning models is that they lack interpretability. Our project aims at integration of the Explainable AI (XAI) with the blockchain technology and determining whether a certain user has a good credit score or not, supplemented with an explanation for the same. In this project a system to generate explanations along with the predictions for every single user and making it securely accessible to everyone is designed. Managing the transparency, security and tracking of decisions made by the model over some time for a user is accomplished by constructing a blockchain.

Project ID 02: 

Project Title : Generation of Deep Fakes Detection using GANS for textual articles.

Student Members: 

Prince Chaudhary, Final Year BTech (IT)

Avani Bajaj, Final Year BTech (IT)

Ajayveer Singh, Final Year BTech (IT)

Vidhi Jain, Final Year BTech (CS)

Objectives and Methodology:

The objective was to create a network that could automatically generate full-length articles and as the fake news publishers have become more sophisticated in their propagation strategies, these machine learning models have been introduced to rapidly detect patterns in news sources and articles. The main task was to provide a classifier that would differentiate between the generated fake text and real text which would help in financial fraud and fake news article detection. 

The technology used is Generative Adversarial Networks (GANs). One network, called the generator, generates new data instances, while the other, the discriminator, evaluates them for authenticity; i.e. the discriminator decides whether each instance of data that it reviews belongs to the actual training dataset or not. Basically we have tried to mislead discriminator using two neural networks, pitting one against the other in order to generate new, synthetic instances of data that can pass for real data. 

Project ID 03: 

Project Title : Segmentation of Blood Vessels from Retinal Fundus Images

Student Members : 

Akash Dholaria, Final Year BTech (CS)

Gauri Nigam, Final Year BTech (CS)

Shambhavi Shikha Tiwari, Final Year BTech (IT)

Rajat Pandey, Final Year BTech (IT)

Objective and Methodology

The primary objective of this project is image segmentation of Retinal Fundus image in order to obtain a pattern of blood vessels, that can be used to classify disease within a retinal image.

To perform segmentation on the biomedical images, we used semantic segmentation because it is a technique that detects, for each pixel, the object category it belongs to, all labels must be known to the model.  In our case, for segmenting an image and classifying it, the input image is a retinal fundus image with the first-hand task of lining and segmenting out the blood vessels pixel by pixel. The architecture used is the U-Net architecture where each pixel is classified to be the part of a vessel or not by predicting a probability and comparing it with a chosen threshold.

Framework of the Methodology

UNet Framework

Project ID 04: 

Project Title: Evaluation of Adversarial training Methods on image based machine learning setups

Students Members: 

Vidit Chokshi, Final Year BTech (CS)

Manthansingh Bisht, Final Year BTech (CS)

Vineeth Nair, Final Year BTech (CS)

Abheek Ranjan Das, Final Year BTech (IT)

Security is a mandatory feature in the field of computer vision. Adversarial attacks are one of the most famous techniques used for breaching security, which attempts to fool AI models by malicious input. For our project, we took the case of chest x-rays, since medical data is extremely fragile and minute tampering can lead to unforeseen and disastrous effects. Especially with the increase of data breaches involving medical data, it has become essential to add new types of security measures to make the models robust. In this project, initially, various types of adversarial attacking strategies are used to introduce noise into the images in the input dataset. These modified images are generally indistinguishable to the human eye, but can easily mislead the model into a completely wrong prediction. The goal is to deceive an already trained model. The defensive strategies help to make the model robust by identifying the attacks and avoid misleading predictions by the model. We implemented three different types of defense strategies against poisoning attacks to check the robustness improvement in the model.

Figure 1: Adversarial machine learning project pipeline

Project ID 05: 

Project Title: Evaluation of Adversarial training Methods on image based machine learning setups

Students Members: 

Vidit Chokshi, Final Year BTech (CS)

Manthansingh Bisht, Final Year BTech (CS)

Vineeth Nair, Final Year BTech (CS)

Abheek Ranjan Das, Final Year BTech (IT)

Security is a mandatory feature in the field of computer vision. Adversarial attacks are one of the most famous techniques used for breaching security, which attempts to fool AI models by malicious input. For our project, we took the case of chest x-rays, since medical data is extremely fragile and minute tampering can lead to unforeseen and disastrous effects. Especially with the increase of data breaches involving medical data, it has become essential to add new types of security measures to make the models robust. In this project, initially, various types of adversarial attacking strategies are used to introduce noise into the images in the input dataset. These modified images are generally indistinguishable to the human eye, but can easily mislead the model into a completely wrong prediction. The goal is to deceive an already trained model. The defensive strategies help to make the model robust by identifying the attacks and avoid misleading predictions by the model. We implemented three different types of defense strategies against poisoning attacks to check the robustness improvement in the model.

Figure 1: Adversarial machine learning project pipeline