Project Showcase
FISAT
CSE Department
Automatic Generation of Web Page from Hand Drawn Sketches
Yedhin Kizhakethara, Ranjith R K
Implementing smooth and engaging experiences for users is a crucial goal for companies of all sizes, and it is a process driven by numerous cycles of prototyping, designing, and user testing. Large organizations and MNCs have the capability and financial power to dedicate entire teams to the design process, which can take several weeks and involve multiple stakeholders. But small startups and businesses do not have these resources, and their user experience may suffer as a result. As webpages are essential for marketing and businesses, there is a need for faster iteration, easier accessibility to non- developers and less dependency on developers for initial prototyping. A deep learning model which can convert hand drawn wireframes of websites into corresponding digital webpages can resolve the mentioned problems.
An Assistive Humanoid Robot for the Elderly
Job Jacob
In the current world, where people strive day and night to meet their day to day goals, people often have to leave their weak and disabled elderly parents alone at home. This situation causes mental distress to both the person staying at home and the people leaving their dear ones alone at home. It will also lead to the need for a personal assistant for the weak or disabled person at home. Hiring a human caregiver is not preferred by many due to the lack of privacy, unskilled employees, and are more often considered an overburden. The need for a robot home assistant is evident in this scenario. The existing robot assistants are highly expensive and are not affordable for the common man. The objective of this project is to develop a cost-effective, reliable, and trustworthy human-like assistive robot. The robot, with cognitive capabili- ties, will be able to verbally and physically interact with humans, send alerts to the necessary people when an emergency arises, and will even provide a ca- pability to be completely monitored and controlled by an authorized remote person. Low-cost, yet strong, building materials, easily available hardware, and openly available software resources will be used in order to develop the robot.
Internet of Things Mobile - Air Pollution Monitoring System
Prasanth P M, Sanjay R, Kevin Jose, Midhun Baby
Air pollution is the largest environmental and public health challenge in the world today. Air pollution leads to adverse effects on Human health, climate and ecosystem. Air is getting polluted because of release of Toxic gases by industries, vehicular emissions and increased concentration of harm- ful gases and particulate matter in the atmosphere.Air pollution affects our day to day activities and quality of life. This project proposes an air pol- lution monitoring system. An IoT module containing gas sensors, mobile application along with a Wi-Fi enabled computing device will be used to monitor the air pollution. The gas sensors can gather data from air and forward the data to the computing device. The computing device can trans- mit the data to the cloud via the Wi-Fi module, also a mobile application is created, that enables the users to access relevant air quality data from the cloud. If the user is traveling to a destination, the pollution level of the entire route can be predicted.Furthermore, air quality data can be used to predict future Air Quality Index (AQI) levels.
Balancing of COVID-19 Dataset Using GAN
Neeraj Krishna M S
Corona viruses are a large family of viruses that cause illness ranging from the common to more deadly diseases such as middle east respiratory syndrome. COVID-19 affects different people in different ways. Although supplies of tests are increasing,it is impossible to test each and every person by a viral test or by using an antibody test. Another simple way to detect the COVID virus is by taking the chest X-ray and analysing it.But the problem is that the number of X-RAY images for analysis is very low and at the same time number of X-RAY images of non-COVID cases is very less compared to the covid cases in the dataset available in GITHUB.So to address this issue I am proposing a solution of generating X-ray images of non-COVID patients with the help of Genative Adversarial Network (GAN) and augment these images with the existing data set to solve the problem of imbalancing nature of the dataset.For the purpose of solving this issue a GAN neural network is implemented. The result of the experiment shows that the proposed method is very much suitable for solving the imbalanced nature of the dataset.
Brainwave Classifier
Nandini Menon
Attention span is the amount of time spent concentrating on a task before becoming distracted. In today’s world when information and distractions constantly surround us, it is becoming increasingly dicult to stay focussed. This project proposes a solution to this problem by creating a system that takes brainwaves as inputs and produces an output of whether or not a person is paying attention.
Super Resolution
Kavya Venugopalan, Mamitha Muraleedharan, Riya V Raj
Super resolution is the process of upscaling or improving the details within an image. Often a low resolution image is taken as an input and the same image is upscaled to a higher resolution, which is the output. The details in the high resolution output are filled in , where the details are es- sentially unknown. Single image super-resolution (SISR), as a fundamental low-level vision problem, has attracted increasing attention in the research community and AI companies. The project aims at analysing and study- ing the three key components of Super-Resolution Generative Adversarial Network (SRGAN), viz., network architecture, adversarial loss and percep- tual loss, and improve each of them to derive an Enhanced SRGAN. The project aims at implementing the concept of Residual-in-Residual Dense Block (RRDB) without batch normalization.
Lane Detection Using Cascadded CNN
Ponnu Varghese, Riya Jose, Silpa M Kuriakose, Swathi S
Lane detection is extremely important for autonomous vehicles. For this reason, many approaches use lane boundary information to locate the vehicle inside the street, or to integrate GPS-based localization. As many other computer vision based tasks, convolutional neural networks (CNNs) represent the state-of-the-art technology to identify lane boundaries. In this work, we present an end-to-end system for lane boundary identification, clustering and classification, based on two cascaded neural networks, that runs in real-time.
Music Genre Classification
Manisha Martin, Rosebelle Joseph, Rosemol Suresh, Sandeep Remesh
Music Genre classification is very useful nowadays due to rapid growth in music tracks, both online and offline. Music genre labels are useful to organize songs, albums, and artists into broader groups that share similar musical characteristics. Music Genre Classification uses different machine learning techniques. Most of the Convolutional Neural Network (CNN) models, and many other tech- niques uses Mel-frequency cepstral coefficients (MFCC) as the feature ex- tracted from the dataset for training and testing the models. Here we try to capture the effect of CNN on the Fast Fourier Transform (FFT) features to provide best model for predicting genres of music.
Image Quilting
Lirine John C, Tania Susan Sajeev, Tara Jose, Teena John
Image quilting is an image-based method for generating visual appear- ance in which a new image is synthesized by stitching together small patches of existing image.Efros and Freeman’s method is a non-parametric patch- based method which takes an input image and computes an output texture image by quilting together the patches.The stitching technique in Efros and Freemans method reduces the transition effect between patches.A complete mathematical description of the linear programming problem is used for quilting and in implementation.
Route Map Creator
Leo Varghese, Mohit Rajan E, Riya Alexander
A Decentralized system to track people’s movement using GPS. This system mainly consists of 3 modules viz; bootstrap server, admin console and a client [Android App]. The client will send a broadcast message to the other clients with the help of the bootstrap server. In the case of a positive result, the other clients are alerted. The location data is shown in the Map and the location history is also shown as a timeline on the client-side.
Content Based Lecture Video Retrieval
Nitha Maria Shaju, Saranya Ashok, Sreelakshmi Pradeep, Varsha Rajan
In recent years the amount of lecture video data on the World Wide Web is growing rapidly. Retrieving necessary videos from a large video archive is a crucial task. Manual retrieval is widely used for retrieving content based video. But, this method is time consuming. So an efficient method can be used to retrieve the necessary video from large video data. Automatic video segmentation and key-frame detection can be applied that is capable of re- trieving content videos that can solve the above mentioned problem. Textual metadata can be extracted by applying video Optical Character Recognition (OCR) technology on key-frames and Automatic Speech Recognition (ASR) on lecture audio tracks.
Data Slicer
Paul Antony, Mohammed Zeeshan, Shreya Kulangara, Zacharia Joshy
Machine learning has become part of different aspects of our life like Healthcare, Transportation, Market and Sales etc. Directly or indirectly Machine learning is influencing our decision-making process. From the field of cancer detection to recommendation systems machine learning has be- come the heart of many systems. A slight increase in the performance of the model can have a huge impact on these systems. The performance of the model is usually affected by class distribution, class imbalance etc of the data it learns from. A small subset of the data might be causing the model to underperform. Identifying and addressing such subsets which lead to the underperformance can enhance the accuracy of the ML model. Manually detecting such subset even with the aid of domain experts is difficult. Due to the sheer volume of data analysing each possible subset becomes an ex- haustive task. It is difficult to address a problematic subset if it is made up of an arbitrary collection of points. Defining a problematic subset in a human-readable form can help in addressing the subset We propose Data Slicer, an automatic data slicing system that finds interpretable subsets of data for which the model under-performs. The system finds the top K problematic slices in the given dataset. It is also able to find subsets in the training data which have influenced the decision- making process of the model and lead to miss prediction. Humans can understand more from visual data than textual data. Data slicer provides a visualization based interface where each slice can be visually analysed.
Automatic Question Tagging
Meleesha Tresa Paul, Pooja Biju, Nousheen Sultana, Swathi Sreekumar
In recent years, computerized adaptive testing (CAT) has gained popularity as an important means to evaluate students’ ability. Assigning tags to test questions is crucial in CAT. Manual tagging is widely used for constructing question banks; however, this approach is time-consuming and might lead to consistency issues. Automatic question tagging, an alternative, has not been studied extensively. The proposal models utilize machine learning method to represent tagged questions along with their keywords.
Hand Gesture Recognition for Assisting Deaf People
Pooja Sivakumar, Raman KP, Neethu S Kumar, Sneha Anna Joshy
Hand gesture is one of the most commonly used method in sign lan- guage for non-verbal communication. It is used by people who have hearing or speech impairments to communicate among themselves and with normal people. Various sign language systems has been developed around the world but they are neither flexible nor cost-effective for the end users.Furthermore there exist minimal dataset for dynamic hand gestures. This work presents a system for real-time gesture prediction with minimal cost and increased flexibility. In the system development process the most optimal classifi- cation algorithm for the application is found. A completely robust hand gesture recognition system is still under heavy research and development. The implemented system serves as an extendable foundation for future work.
Augmented Reality 3D Model Visualization App for Learning
Jyothilakshmi M.J, Krishnaja Raju, Sarah Varghese, Varsha George
Augmented Reality(AR) has huge potential in aiding students to master concepts taught in class. In the classroom, the problem becomes particularly relevant when trying to explain to students using 3D shapes. AR technology has ability to render objects that are hard to imagine and turn them into 3D models, thus making it easier to comprehend. The use of only 2D images to navigate a 3D object can be problematic. An AR based tool was developed for better understanding of complex 2D images superimposed on textbooks.
Automatic Extraction of Access Control Policies from Natural Language Documents
Sreerag O.B, Sayooj A.P, Nandu V.K, Yadhukrishnan P.C
Almost all applications that deal with safety, privacy or defense include some form of access control.However, the initial development of Access Con- trol Policies (ACPs) can be very challenging, it is time consuming.Here, a new framework towards extracting ACPs from unrestricted natural language documents is proposed.It is done using a deep learning model and Visualized using dependency parsing tree. The aim is to correctly identify the Access Control Policies (ACP) elements, which describes allowable operations of the system, from Natural language documents using NLP.
Information Extraction and Inferencing System for Substance Abuse Cases
Sanjay Govind P, Tibi Sunny, Shijas Muhammed T P, Judith George Joseph
The rising number of substance abuse cases is a serious situation that demands significant attention from authorities. A system that can provide valuable information regarding this epidemic can be helpful in handling these issues efficiently. The developed system extract the news reported on substance abuse related crimes from online resources like online news pa- pers, articles etc. Then process the extracted data using Natural Language Processing techniques to generate a set of information that can be helpful in generating valuable inferences.
Large Scale Scene Classifier
Pooja Jayakumar, Rinu Rahim, Risana Rahim, Saira Sajeer
Scene categories are often defined according to its functions and there exist large intra-class variations in a single scene category. Deep neural net- works demonstrate to have a high performance on image classification tasks while being more difficult to train. Due to the complexity and vanishing gradient problem, it normally takes a lot of time and more computational power to train deeper neural networks. Deep residual networks (ResNets) can make the training process faster and attain more accuracy compared to their equivalent neural networks The model focuses on large-scale scene recognition using a fine-tuned ResNet50 on the Places365-Standard dataset and compared it with the other trained-from-scratch Places-CNNs for scene classification. The ResNet eases the optimization by providing faster convergence at the early stage.