Research and Development (2017-2022)

Unsupervised learning

Motivated by the fact that the conventional clustering algorithms such as K-Means, DBSCAN, etc. are not suitable for handling big data, two different approaches have been adopted to develop such algorithms. In the first approach, clustering is performed in a sampled data set and also by checking the stability of cluster centroids by repeated k-Means the cluster number is selected. The algorithm is called Cluster Number Assisted k-Means (CNAK) (Saha and Mukherjee’ 2021, PR). It is also observed the hierarchy of clusters at different values of k could also be determined by this process.

In the second approach, a multilayer feedforward network architecture is used to perform clustering by minimising an entropy based loss function, which has two components, classification entropy and class entropy. The loss function is called RECAL loss (Saha and Mukherjee; 2019, NCVPRIPG, and Chatterjee et al’ 2022, PRL). The concept is shown to be useful for deep networks and suitable for handling large datasets.

Related Publications

 

Model compression

The objectives set here are to develop techniques for deriving equivalent deep neural models of smaller size to make them suitable for running in low resource computing platforms. Apart from following conventional approach of tensor decomposition and running equivalent filters with a cascaded set of low-ranked filters (Saha et al.’ 2019, ICIP) and knowledge distillation (Sadhukhan et al’ 2020, ICIP), we have also considered hierarchical decomposition of the network and breaking a generalised large model into subnetworks handling classes in a smaller domain (Sairam et al’2018, ICVGIP, Sadhukhan et al’ 2022, ICIP).

Related Publications

Computer Vision  

A bi-LSTM based deep learning architecture has been proposed for video summarization (Dutt et al, ICIP’2018). A novel architecture (Saha et al., ICIP’2018) combining residual network and LSTM followed by an attentive processing has been proposed for extracting features for its application in comparing a pair of videos for the task of person re-identification. Analyzing eyegazes of a viewer from a wearable video eyetracking device, an algorithm for video retargeting has been developed, which also takes care of the problem of stabilization of the egocentric video (Bhattacharya et al, ICIP’2022).

New algorithms for estimating depth and disparity (Nehra et al, ICVGIP’2021) and performing super-resolution (Kar et al, ICIP’2022) of images of a light field camera have been developed.

Related Publications

  Remote Sensing image processing  

In this area, various techniques for processing both multispectral and SAR images have been developed for land cover analysis, and studying various events and environmental issues of mining coals, etc.

A simple set of rules using spectral slope have been proposed to use to classify pixels in multispectral Landsat images into three broad categories such as Waterbody, Vegetation and Vegetation Void classes. A semi-supervised technique has been proposed to perform land cover analysis under these broad categories and further using unsupervised techniques such as k-Means clustering more subcategories are discovered (Shashaank et al’ ICAPR’2017). The technique is made more robust and discriminatory by using Hybrid Pol SAR images of the same region providing finer categories of forestry (dense and shallow), vegetation and water bodies (Shashaank et al’ IJRS’2020, Shashaank et al ‘ IGARSS, 2019). The multimodal analysis has become useful also to detect surface mines (Shashaank et al’ ICVGIP’ 2018). Using features from trained supervised networks, unsupervised landcover analysis (Saha et al’ IGARSS’ 2020) has also been performed by optimizing RECAL loss function (Saha et al’ NCVPRIPG’2019). A novel transformation invariant algorithm has also been proposed for developing a model of supervised classification of hyper spectral images (Saha et al’ IGARSS’2021).

For studying various events and environmental issues of coal mines multispectral Landsat images are used. Observing seasonal behaviour of temporal changes, in particular, surface temperature, subsurface fire in coal mine areas have been detected (Mukherjee et al, NCVPRIPG’2017). A novel index for detection of coal mine has been proposed (Mukherjee et al, IEEE JoSTAER’ 2019a). Using it various phenomena such as separation of mine and non-mine water in the coal mine area (Mukherjee et al, IEEE JoSTAER’ 2019b, NCVPRIPG’ 2019 ), open surface mine (Mukherjee et al, IGARSS’2019), detection of activities related to reclamation of mine areas (Mukherjee et al, PREMI’2019), detection of coal overburden region (Mukherjee et al, MMTP’ 2021), and separation of various types of regions in surface coal mine area (Mukherjee et al, IGARSS’2022).

For characterising dual pol SAR images a deep clustering algorithm (Chatterjee et al’ IEEE GRSL’2021) has been developed, which also uses the RECAL loss function (Saha et al’ NCVPRIPG’2019). The technique has been applied for detecting paddy fields across various seasons of cultivation in Eastern India (Chatterjee et al, IJRS’2021). In another interesting study, quad pol SAR images from the Indian Lunar Mission Chandrayaan-2, has been used to study the properties of lunar impact craters with the objective of determining whether they could possibly contain ice (Putrevu et al’ LPSC’2022).

Related Publications

Medical Image Processing  

Novel techniques for processing 3-D medical images in the modalities of CBCT and MRI have been developed. For treatment planning of colorectal patients, an algorithm for more accurate estimation of margins of radiotherapy has been developed by studying translational and rotational shifts of pelvic structures. For this automated detection of a few reference anatomical part is developed (Malladi et al, PREMI’ 2017). A robust registration algorithm for this purpose has been developed using multiple estimates of randomly sampled set of points (Malladi et al, PRL’2018). A technique for segmenting tumours in CBCT images of lungs has been developed (Veduruparthi et al, ICIP’ 2018) by applying concepts of level sets and local rank transform (LRT). A deep neural architecture employing spatial attention network (Mazumder’ Neurocomp’2022) has been found to be effective and efficient (of lighter size compared to SOTA models) for segmenting brain tumours in MRI images.

Related Publications

Medical Informatics and Telemedicine System 

A new version of the telemedicine system iMediX (Mukhopadhyay’ HMS: Taylor & Francis’2020) has been developed to blend home care with hospital care during the first few months of COVID pandemic. The system had been used to treat residential patients at IIT Kharagpur during the pandemic and is being presently integrated in the Hospital Management System of the BC Roy Technology, Hospital, IIT Kharagpur. Another cloud based version of the system (called Vivek Sanjivani) has been installed at Ramakrishna Mission Home of Services, Varanasi facilitating their community outreach health care program for twelve remote centers, including six mobile centers in April’2021. During 2022-23, Vivek Sanjivani served 27,875 patients through Stationary Telemedicine Units and Mobile Telemedicine Units and conducted Non-Communicable Disease Screening for 18,878 patients (https://www.iitkgpalumnifoundation.in/newsroom/news/IIT-Kharagpur-develops-Matri-Seva-a-health-database-free-software-for-Varanasis-RK-Mission-Home-of-Services.dz). An open sourced version of the system called iMediXcare (https://github.com/jmGithub2021/iMediXcare) has been released in the public domain.

In collaboration with Tata Medical Center, Kolkata, a medical image data bank, called CHAVI (https://chavi.ai) for archiving radiological images of cancer patients has been developed (Kundu et al, JDI’2022). A new methodology of deidentification of patient’s clinical data and images has been developed (Kundu et al., JMS’2020a and JMS’2020b). The process flexibly takes care of requirements of information modeling of various studies and subsequently mapping them to a relational database schema (Kundu et al’ JDI’2021).

In another collaboration with Tata Translational Cancer Research Centre, Tata Medical Center, Tata Medical Center, Kolkata, a study on adherence of maintenance therapy of paediatric patients suffering from Acute Lymphoblastic Leukaemia (ALL) has been carried out and it is shown that there is a need to increase awareness and clarity in treatment protocol among the doctors (Mungle’JPHO’2020).

For assessing the level of understanding and the requirements of community health worker engaged in performing prenatal and antenatal checks, a survey has been conducted (Ghosh et al, CEGH’2021).

In another work, implementing a knowledge graph on the interactions among drugs, diseases and genes, a system has been developed (Saha et al, MBEC’2020). The system integrates information from various types of publicly available repositories and is capable of performing complex queries providing valuable insight from the integrated information system.

Related Publications

Cancer radiomics  

Extracting radiomic features from 18 FDG PET/CT images of patients suffering from head and neck cancer, machine learning based models for predicting their survival outcome have been developed (Ghosh et al, CMPB’2020). Using the deidentified data from CHAVI medical image data bank two studies were carried on for predicting outcome of patients suffering from lung cancer (Das et al., CMPBU’2022) and on studying association among molecular profile and changes in radiomic features in MRIs of patients suffering from brain tumour and subjected to chemoradiation (Achari et al, RO’2021).

Related Publications

Document Image Processing 

A dataset has been released for treating margin noise of documents (Dey et al., OST@ICDAR’2017). A novel technique on detection and localization of struck out words in hand written documents has been developed (Poddar et al, ICDAR Workshop’s 2021). In another deep learning based technique under the generative adversarial learning framework for restoring struck out and underlined words handwritten has been developed (Poddar et al, ICVGIP’2021).

Related Publications

  Affective computing

Various algorithms analyzing eye gaze data to study human psycho-visual cognition process have been developed. In one such study, spatial distribution of fixation points of eyegazes are used to predict the viewer’s interpretation of ambiguous images (Roy et al, IEEE TAC’2017) . Eyegazes of viewers’ participating in Rorschach Ink Blot test are also studied to relate with normative findings of standard questionnaire based studies. Video eye tracking data has also been used to develop computational models for characterizing eyegazes with the content of a video (Malladi et al, MMSP’2020).

In another study, a mobile based app called MonTaj, has been developed to assess the level of attention and working memory of young adults through their participation in specially designed computer games (Basu et al, JAIHC’2022). A mobile app (Ansari et al., ICALT’2020) has also been developed to enhance reading a capacity of dyslexic learners.

For studying and developing resilience among students a system has been developed in consultation with the Students’ Counselling Center, IIT Kharagpur (Roy et al., COMSNETS’ 2017).

Related Publications

Robotic path planning and tracking  

Algorithms on trajectory planning under various scenario of robotic movement have been considered. In robot soccer (for Small Scale League (SSL) competitions), an overhead camera captures the 2-D coordinate positions of robots and the soccer ball in a field. Two novel algorithms exploiting priors (Agarwalla et al, COMADS/CODS’ 2018) and adapting concept of potential field in the deep learning framework (Koukuntla et al, CACRE’ 2019) have been proposed. In another scenario for indoor (Bhowmick et al, ICVGIP’ 2018) and outdoor (Bhowmick et al, IEEE SSCI’ 2018) movement of a robot using a monocular camera an algorithm for learning topological map and routing tables for navigation have been developed. In another approach, correlation filter based search among multipath trajectories is carried out (Bhunia et al’ PREMI’ 2019, CVIP’2020).

Related Publications

Computational Phylogenetics  

Continuing our previous work on development of algorithms for building super-trees from smaller phylogenetic trees, a novel algorithm for inferring species tree using internode distances and excess gene leaf count has been proposed (Bhattacharyya and Mukherjee’ JME 2017).

For comparing phylogenetic trees a new measure, called deformity index, has been proposed, which is found to be more discriminatory than existing measures (Mahapatra and Mukherjee’ JME, 2021).

A novel large scale genomic feature called Genomic Foot Print (GFP) has been proposed to perform taxonomy based classification among closely related species (Mahapatra and Mukherjee’ PREMI, 2019). The method has been found to be more suitable for building large phylogenetic trees from mitochondrial sequences of different species (Mahapatra and Mukherjee’ BIBM, 2020; CCHTS’2022). An interesting study on prehistoric human origin and migration has also been reported (Mahapatra and Mukherjee’ bioRxiv, 2020).

Related Publications

Visual Saliency 

A free viewing eye tracking data set with egocentric videos has been released in the public domain (Malladi et al, IEEE Access’ 2022). The dataset is expected to contribute in advancement of research in visual saliency for videos and affective computing on videos.

A lighter deep neural model has been proposed to compute salient regions in images (Malladi et al, ICIP’2021). The model has been further extended to videos (Malladi et al., ICIP’2022). A technique to draw explanation from a deep visual saliency model from the view point of classical Gabor filter based models has also been proposed (Malladi et al. , CVIU’2023).

Related Publications

Compressed domain processing  

Extending the theory developed for separable filters previously, a new algorithm for non-separable filters in the block DCT domain also has been proposed (Mukhopadhyay and Sairam, ICIP’2018).

Algorithms for resizing and color enhancement in the block DCT domain are revisited and refined (Mukhopadhyay’ ISCAS’2017; Mukhopadhyay’ ISSCS’2017).

Related Publications

Digital Geometry  

New algorithms for efficiently computing of inter-simplex Chebyshev distances among 2-manifold surfaces for voxelization have been developed. (Bhunre et al’ IS’ 2018; IWCIA’ 2017). By computing inner and outer isothetic covers of characters a rough set theoretic approach based algorithm has been developed for font independent character recognition (Chaudhuri et al., PREMI’ 2017; ICAPR’ 2017). In a research monograph, different classes of digital distances and their roles in approximating Euclidean metric have been studied (Mukhopadhyay’ Elsevier’ 2020).

Related Publications

Biomedical instrumentation  

A novel sensing technique for performing bed side impedance cardiography has been developed (Ghosh et al’ Springer’ 2022). The instrument has been used for measuring blood pressure (Ghosh et al’ IM’ 2022), estimating stroke volume and ejection fraction (Ghosh et al, EMBC’ 2018), and various other measures and parameters those conventionally obtained from echocardiograms (Ghosh et al, AIM’ 2019). The technique has been found to be effective in detecting and localizing lesions in coronary artery. The above research has been carried out in collaboration with the Department of Cardiology, Kolkata Medical College. The data set has been captured from cardiac patients of the same Department and the validation and testing have been carried out by the doctors there.

In another approach, photo plethysmograph (PPG) signals are analyzed to measure blood pressure with high accuracy non-invasively (Chakraborty et al, TIM’ 2021, Chakraborty et al, TCE' 2023).

An image processing based system for measuring haemoglobin from images of blood drops on a filter paper has been developed (Ghosh et al’ CMPB’ 2023). The system has been found to useful to detect anaemia among pregnant women.

Related Publications