Department of Computer Science and Engineering

Permanent URI for this collection

http://nits.ndl.gov.in/handle/123456789/111

Browse

Now showing 1 - 20 of 31

A Comparative Study of Sentiment and Emotional Analysis of Parallel Corpora (Manipuri-English)
(National Institute of Technology, Silchar, 2026) Devi, Thoudam Siyabeti
A Hybrid Model for Analysis and Detection of Covid 19 on X Ray Images using Machine Learning and Deep learning Techniques
(Department of Computer Science & Engineering, NITS, 2022) Varma, Vivek
Covid-19 has presented itself as a deadly infection that can affect all sorts of living beings. The day-to-day life of humans, in all of the countries are getting hindered due to this rapidly spreading viral problem. A clinically based examination of COVID-19 affected persons revealed that most of these people had a respiratory based infection after coming into touch with this disease. Chest based x-ray and chest Computer Tomography are better effective imaging techniques for finding respiratory problems. Nonetheless, producing a x-ray is a lower cost procedure than a CT scan. Deep learning is a very successful machine learning procedure, providing valuable perspective to examine a large quantity of x-ray images, which may provide a significant impact on Coronavirus screening. This work takes x-ray images of chest from corona infected and healthy people. The images are balanced and then fed into convolutional neural network models and their performance is analysed. The accuracy of ResNeXt, Xception and InceptionV3 is checked. 15,135 x-ray scans of chest were taken as the dataset which is collected from multiple certified sources, from this 12,121 are used in training and 3032 are used in validation. The Xception model provides more accuracy (i.e., 97%) on detecting covid in the images outperforming other pre-trained models. Furthermore a concatenated model consisting of all the 3 above stated models was implemented and trained on the same data set which provided the highest overall accuracy of 98%. A second instance of the same concatenated model was created and trained on a new dataset collected from Brixia project consisting of only covid x-rays divided into 18 classes on the basis of their severity. This model was then used to assign a severity score on the covid x-ray scans providing 92% accuracy. Thus the final model being proposed is able to take the x-rays and classify them into three categories (Pneumonia, Normal, Covid) and assign a severity score to it in case of detection of Covid.
A Multimodal Biometric Recognition System Based on Palmprint and Dorsal Hand Vein Using DenseNet
(National Institute of Technology, Silchar, 2026) Priyanka
Multimodal biometric systems have emerged as an effective solution to overcome the limitations of unimodal systems, such as susceptibility to noise, spoofing, and limited discriminative capability. Among various biometric traits, palmprint and dorsal hand vein modalities provide high distinctiveness and reliability. In particular, dorsal hand vein patterns, being subcutaneous, offer enhanced resistance to forgery and environmental variations. This thesis is a multimodal biometric recognition system that implement palm print and dorsal hand vein images using deep learning based feature ex- traction techniques. Different fusion Methods such as feature, score, and hybrid fusion, are evaluated to determine the best approach. Experimental results shows that hybrid fusion achieves best performance by effectively combining score and feature fusion. Furthermore, a Siamese DenseNet model with triplet loss and arcFace loss is employed to learn discriminative embeddings and acheived an Area Under Curve (AUC) of 0.991 and an Equal Error Rate (EER) of 0.045%. The results highlight the effectiveness of combining multiple biometric modalities with advanced fusion techniques for robust authentication.
A Novel Automatic Shot Boundary Detection Algorithm Robust to Illumination and Motion Effect
(Department of Computer Science & Engineering, NITS, 2019) Singh, Alok
Apparently, the summarizing, classifying and indexing valuable information from the massive amount of videos is very inefficient and time consuming through manual procedure. In order to make this process effective, it is necessary to segment a video into meaningful shots and locate the boundaries between the shots. Shot is a collection of continuous frames having similar content and visual feature and when the continuity in the similarity of the frames breaks it is termed as transition from one shot to another shot which creates a boundary between the shots. But, the performance of shot boundary detection approaches are yet to be addressed for the videos having sudden flash light and Object/Camera Motion efficiently. In this thesis, a novel approach for abrupt transition detection is proposed which can able to withstand under certain illumination and motion effect. The proposed algorithm can be roughly divided into two stages: firstly, the adaptive Wiener filter is applied to the lightness component of the frame to retain some important information on both frequencies and LBP-HF is extracted to reduce the illumination effect. From the experimentation, it is also confirmed that the motion effect is also reduced in the first stage. Secondly, Canny edge difference is used to further remove the illumination and motion effect which is not handled in the first stage. TRECVid 2001 and 2007 datasets are applied to analyze and validate our proposed algorithm. Experimental results manifest that the proposed system outperform the state-of-art shot boundary detection techniques.
A Sarcasm-aware Cyberbully Detection Framework for Code-Mixed Hinglish Text Using Deep Learning
(National Institute of Technology, Silchar, 2026) Rai, Sweety
Cyberbullying detection in Hinglish social media text is challenging due to code-mixing, informal language, and sarcasm that often masks harmful intent. Traditional text-based approaches struggle to capture such implicit expressions. This research pro-poses a two-phase sarcasm-aware framework for cyberbullying detection. In Phase 1, a MuRIL-based model is trained to learn sarcasm-aware representations from Hinglish cyberbullying-related data. In Phase 2, these representations are integrated with tex-tual and emoji signals for classification. An experimental setup involving model variants indicates the consistent improvement, with the proposed model achieving F1-score of 85.70%, in comparison to the text-only baseline which achieves 83.69%. Stratified analysis further shows improved performance on sarcastic samples, reducing the sarcasm gap from 27.19% to 18.83%. Cross-dataset evaluation on Mendeley and TRAC datasets shows improvements in AUC and overall F1-score. However, the observed F1 score im-provement is mainly due to better detection of non-aggressive samples, highlighting the effect of domain mismatch on aggressive classes. These findings highlight the im-portance of sarcasm-aware modelling while also reveals the limitations in cross-domain generalization.
A Smart Farming Model for Crop Yield Prediction Using Machine Learning Analytics
(National Institute of Technology, Silchar, 2026) Saha, Sukanya
A Symmetric Cryptography Algorithm with 1024 bit Data Block and 512 bit Key with Enhanced Fiestel Cipher
(Department of Computer Science & Engineering, NITS, 2020) Maheshwari, Bharat
The growth of latest technology specially internet and its applications imposes a major concern in respect to information security in terms of protecting essential and confidential content against any possible and vital attacks. Cryptography has made the difference over the period of time by using developed encryption algorithms where some of them have successfully fulfil its designed goals, however others have failed due to specific requirements or encryption/decryption procedures have been recognized. Recent encryption algorithms employed aims mainly to escalate the security levels of sensitive information in order to deliver confidentiality, authentication, integrity, and non-repudiation. In this study, the proposed encryption algorithm has 1024 bit data block as input and combined the permutation and substitution methods for in-round encryption / decryption in-side the enhanced fiestel cipher along with pseudo randomly generated 512 bit key. Key generation and encryption / decryption performing using pseudo random number generation technique and enhanced fiestel cipher with 7 rounds methodology which removes of previous vulnerabilities. Analysis and experimental results have clearly proven solid randomness on key generation, encryption and decryption also enhanced performance and security in which it has met all desired objectives.
Abstractive Text Sumarization Evaluation
(Department of Computer Science & Engineering, NITS, 2022) Tawmo
Abstractive Text Summarization is a burgeoning natural language processing task that has seen with the deep learning models. With the soaring amassment Of un- structured text data, text summarization has found applications in many domains such as law, medicine and news. The growth in research interest in developing better sum- marization systems, more fluent human-like summaries, and significant, more refined datasets have offered to test natural language processing models to improve existing model performance. Even so, no significant improvements are made while evaluating the machine summaries, still relying on metrics such as BLEU and ROUGE which only performs syntactic evaluation and cannot evaluate based On semantic equivalence. Semantic Textual Similarity is a Natural Language Processing (NLP) task that uses a predetermined metric to score the relationship between texts or documents. Measuring semantic similarity scores between texts is useful in many applications in Natural Language Processing (NLP) and fields. In this work, we propose a novel evaluation method to semantically analyse and measure the quality of the abstractive text summaries. The proposed model compared the resultant output across the datasets to determine the proficiency of the model and the datasets with regards to ROUGE, BLEU scores and the baseline system.
Active Dataset Adaptation Based Covid-19 Detection via Deep learning Techniques
(Department of Computer Science & Engineering, NITS, 2022) Gaikwad, Sushant Suresh
An epidemic of Covid-19 disease resulting from the 2019 novel coronavirus started in Wuhan, China, at the end of 2019 and has outspread worldwide. Till now many Covid-19 detection models using Deep learning models has and these models are giving promising lab results but most of them are trained and tested On same Of data only, but when it to real world scenarios, these models are overfitting and give less accuracy because of image parameter differences in lab images and real world images. This work mainly deals with resolving the overfitting problem generated after testing the model on new image data. A single or homogeneous dataset with different sources is used for the experiment, which has somewhat similar images but belongs from different sources. In initial experimental results, the base CNN model gives 70.06 % training and 64.23 % validation accuracy. Whereas VGG 16, training and validation accuracy is better than base CNN with 81.54 %, and 75.62 % respectively. Finally the Resnet50 model giving training and validation accuracy is somewhat improved than VGG16 and much more improved than base CNN with 90.23 %, and 85.56 % respectively. These models are used for testing new images from a different source. On new data, the CNN has a accuracy of 55.23 VGG16 has a test accuracy of 68.34 and Resnet50 has a test accuracy of 77.32 Vo. To overcome the overfitting problem on new image data, all models are again with new data images using an active learning approach. The Least confidence technique is to calculate uncertain images to retrain. After active learning training with three cycles, it is found that test accuracy is somewhat increased on new data for all models but not increased to previously trained data accuracy. The new test accuracy for CNN, VGG16, and Resnet50 arc 60.53, 71.94, and 80.65, respectively. Resnet50 is giving the best in all cases among all implemented models.
An Automated Biometric System for Overlapped Fingerprint Detection and Recognition
(National Institute of Technology, Silchar, 2026) Panigrahi, Srilipta
Recognition of overlapped fingerprint is a non-trivial problem due to complex mixing of ridge patterns from multiple fingers. Most of the existing approaches focus on separation, with less work on a fully automated, complete separation and matching pipeline. In our work, we have proposed a fully automated framework for overlap detection, separation, and identification of constituent fingerprints. A large dataset of overlapped fingerprints is generated synthetically to train SSD model for accurate overlap localization. For separation of constituent fingerprints, an ellipse based overlap removal stategy followed by an orientation-guided spectral clustering approach has been implemented. Further, fusion of similarity scores from SIFT and wavelet features is used for the recognition of separated fingerprints.
Concealed object detection through customized YOLO model
(Department of Computer Science & Engineering, NITS, 2022) Annadurai, Karthic
Concealed Object detection (COD) is a primary need for the automatic security surveil lance system. To detect those objects, first, we need to acquire high-quality data for training the model. Here Terahertz image is being used, which is captured in the fre quency range of 140 GHz that will not cause any harm to humans. Even pregnant women and heart patients can walk through the Terahertz image sensing equipment without any hesitation. Accuracy of the COD is analyzed by training various YOLO models with the customized parameters and the Terahertz Imaging dataset. Various methodologies and analyses have been proposed in this project. The analysis report will be useful for further research activities. YOLO (You Only Look Once) version 5 uses a non-max suppression algorithm by default to calculate the bounding box of the concealed objects. To enhance the accuracy, an ensemble technique available called Weighted Boxes Fusion (WBF) can be implemented. To overcome the data overfitting or less amount of dataset, the augmentation techniques like any adversarial generative neural network can be applied. Labeling those augmented images are mandatory for training any YOLO models.
Construction of Discrete Medial Axis using Digital Geometric Approach
(Department of Computer Science & Engineering, NITS, 2019) Dutta, Dona
Medial Axis is one of the classical problems in the field of computational geometry. The medial axis of an object is defined as the set of all points that has two or more than two closest points on the boundary of the object. It was first introduced by Blum as a tool for biological shape recognition. It is used for shape recognition, character recognition, pattern recognition, geographic information systems(GIS), mesh generation, detail suppression and model simplification for engineering analysis. But most of the existing algorithms for me dial axis construction is for the real domain. For applications in the discrete domain that require medial axis transformation, the medial axis is first constructed using real geometric techniques and then discretized to obtain the Discrete Medial Axis (DMA). Such DMA are often less accurate and can also be disconnected at places. The problems associated with such existing algorithms mainly includes distance error, hardware-precision error, complex algebraic and trignometric functions, resolution error, generation of debris pixels and spurious neighbors. To avoid such discripencies arising due to two-fold domain change error, we propose an algorithm to construct DMA using digital-geometric techniques only. Our proposed algorithm eliminates all the errors present in the existing algorithms. The proposed algorithm constructs DMA of given convex polygons, concave polygons and curved-shape objects by using Discrete Voronoi Diagram(DVD) for extended sites followed by edge detection and edge elimination wherever necessary. Experimental results show the efficacy of our proposed algorithm.
Discrete medial axis extraction using discrete voronoi diagram
(Department of Computer Science & Engineering, NITS, 2022) Agre, Ankit
Voronoi diagram is one of the classical problems in Computational Geometry. Given a set of sites it partitions the plane into subdivisions or regions corresponding to each site such that any point in a given region is nearest to the (Xn-responding site contained in that region. It is One Of the few interdisciplinary concepts with relevant material to found in, but not limited to, astronomy, biology, geography, medical sciences, urban and regional planning so researchers have tried to develop algorithm for computing Voronoi diagram the middle of 20th century. The most common algorithm available in the literature is the Incremental algorithm which is of O(n2) complexity and relatively easy to implement, the Divide and Conquer approach with O(nlogn) complexity and the Fortune's algorithm with complexity O(nlogn) and is by far the optimal solution. But the main problem with all above mentioned algorithms is that they fail to perform with the same accuracy when on discrete space This is because when we compute a Discrete Voronoi diagram we have to consider a finite domain (reptBentecl as a uniform grid) such that the Discrete Voronoi region of site is the finite set of cells closer to that site than any other. Thus, to solve the problem we have a parallel algorithm for the construction of discrete voronoi diagram. Moreover, we have also proposed an algorithm for the extraction of discrete medial axis based on the discrete voronoi diagram. We have also mentioned the applicability of discrete axis for the efficient reconstruction of 3D surface.
Dynamic Prediction Model For Mutual Fund NAV During Trading Hours
(National Institute of Technology, Silchar, 2026) Anand, Mukul
Efficacy of Recurrent Neural Networks in rainfall forecasting
(Department of Computer Science & Engineering, NITS, 2022) Mandal, Sudipta
India is an agriculture dependent country. Agriculture is the primary source of income for maximum Of Indians for survival as well as to use crops as an export material. TO achieve a good agricultural profit a proper knowledge of the trend of rainfall is needed so that crop can be saved from flood and drought situation. So predicting rainfall has been a major task in research works. Rainfall prediction provides an awareness to humans and know rainfall amount beforehand so that they can take certain precautions to save crops from rainfall. Many approaches has come into subsistence in order to predict rainfall. Statistical models, Numerical weather prediction models were used initially for prediction. The statistical techniques and numerical methods for rainfall were unable to show well performance for long term rainfall forecasting due to the dynamicity of rainfall. Due to the non-linear and dynamic behaviour, advanced computer modelling is required to predict rainfall accurately, in lieu Of unreliable statistical techniques. In recent days, Machine Learning (ML) such as Artificial Neural Network (ANN) and Deep Learning (DL) enables self-learning to design a time series, data-driven model on rainfall dataset. This work proposes a ML model Short term Data Driven Rainfall Forecasting Model (SDDRF), which uses Long Short Term Memory (LSTM)-based rainfall prediction model to forecast weather and rainfall in the region Of Shillong , a city in Meghalaya. This thesis considers rainfall prediction using LSTM and is compared with three other neural network models. The study intends to the efficacy of these architectures in rainfall prediction, and determines the best among all, for future application and analysis.
Efficient Fall Detection Technique Based on Markov Model: A Multi-Feature Analysis Approach
(National Institute of Technology, Silchar, 2026) Gaur, Ankur Kumar
Fall detection plays an important role in healthcare monitoring, particularly for elderly patients and individuals because it triggers an alert and allows for faster interventions when a patient experiences a fall. This work presents a lightweight, vision-based fall detection framework designed for real-time deployment in monitoring environments. We propose a Hidden Markov Model-based approach that leverages YOLOv8 pose es-timation to detect fall events from video streams. Our method extracts multi-feature from human pose key point, including head position, body angle, center of mass, ve-locity etc. These features capture various fall patterns. The HMM model providing robust detection and achieve competitive accuracy while meeting real-time processing requirements, making it suitable for continuous monitoring applications.
Fake News Detection of Hindi News Dataset using various Machine Learning Based Approaches
(Department of Computer Science & Engineering, NITS, 2022) Kumar, Sudhanshu
Fake news has been in existence even before the rise of the printing press. But, the rise in the number of fake news has been exponential in the age of social media and the internet. It has become very easy for anyone to post, comment, or share whatever he/she wants with the help of a device and the internet. Detecting fake news is difficult because the content is made to look so reliable that it seems to be coming from a reliable source. Although the problem of fake news is existing for centuries, it is become more dangerous because of the internet and social media. Mostly false information is spread through social media despite lots of investment are done by the social media companies to find fake news and remove them. Although lots of research have been done in this field of fake news, very less work has been reported for the Hindi language. As such, most of the research work is biased about the news of a particular domain (i.e. politics) which may cause in dataset bias. For our work, a Hindi news dataset is created. For feature extraction, Word2Vec and tf-idf have been used. For the classification task, 2 traditional machine learning approaches, i.e. Naive Bayes and Logistic regression, 1 deep learning approach i.e. LSTM and 2 advanced pre-trained approach are used based on BERT and RoBERTa. Based on overall accuracy all over the datasets, logistic regression outperforms the Naive Bayes model evaluated using a conventional ML approach. The efficiency of the LSTM model is 94.45 percent. BERT, out of the two pre-trained sophisticated NLP models we worked on (i.e., BERT and RoBERTa), has the best result of 98.13 percent.
Feature Selection Algorithm Using Effective Range and Overlapping Area for Classification of Gene Expression Data
(Department of Computer Science & Engineering, NITS, 2019) Verma, Manish Kumar
Feature selection is an important issue in the field of machine learning and data mining. Classifier’s accuracies and other generalization performance depend on the outcome of the feature selection. Feature selection can help in building a simple model with high accuracy to solve a complex real-world problem. Prediction and diagnosis of disease at an early stage can help many people to get rid of that disease. Classification of gene expression data can play a major role in the prediction of disease. Recently statistically defined feature selection algorithm named Effective Range based Gene Selection (ERGS) and Improved Feature Selection based on Effective Range (IFSER) is proposed. Where, ERGS considers only the overlapping area (OA) of the effective range of each class for every feature, on the other hand, IFSER also uses the other relationship (inclusion relation) of the effective range of each class. But both ERGS and IFSER fails to handle the one situation when the effective range of one class is completely included in the effective range of the other class. And the overlapping area and number of samples in the overlapping area and including area is incorrectly measures by both the algorithms. In this thesis, we are presenting a modified algorithm for feature selection which will overcome the limitation of ERGS and IFSER algorithms. To showcase the effectiveness of our proposed method several experiments on different gene expression datasets are performed and results are compared with other well-known feature selection algorithms. Experimental results manifest that the proposed modification in ERGS and IFSER improves the classification accuracies on gene expression dataset.
Hybrid Approach to Improve the performance of Ensemble Techniques in Financial Fraud Detection
(Department of Computer Science & Engineering, NITS, 2022) Borah, Abhigyan
In recent times, the use of payment mediums like credit cards have increased and thus the number of fraudulent transactions as well. In these situations where anomaly detection is very important, an unbalanced data-set can lead to a drastic fall in the performance of the classification algorithms. If the majority instances outnumber the number of instances of the minority class it is known as Class Imbalance. Methods like ensemble techniques and sampling techniques have been proposed to handle class imbalance. The ensemble techniques with and without sampling have been executed and it was observed that with the traditional ensemble techniques to detect credit card frauds, the performance of some of the algorithms is considerably low as compared to others. Thus, with the aim to improve the performances of those algorithms, a hybrid approach has been proposed, where two of the worst-performing ensemble techniques have been taken based on their F1 scores. This model takes the output of these two ensemble techniques and predicts the final output based on a logical connector. It was observed that in some of the cases, the improvement in the F1 score is as high as 15.36%, whereas in some other cases the improvements are very minute.
Hybrid MuRIL–CNN Suffix Features for Stable Sanskrit Subanta–Tiṅanta–Avyaya Token Classification
(National Institute of Technology, Silchar, 2026) Kumar, Deepu
Sanskrit, one of the oldest Indo-European languages, presents unique challenges for natural language processing due to its exceptionally intricate morphological system governed by Pāṇini’s grammar. The accurate classification of Sanskrit word forms into subanta (सुबन्त, nominal), tiṅanta (ितङन्त, verbal), and avyaya (अव्यय, indeclinable) categories serves as a prerequisite for a wide range of downstream tasks, including pars-ing, translation, morphological analysis, and sandhi resolution. Despite the growing availability of annotated Sanskrit corpora, computational processing of the language remains challenging due to its low-resource status, agglutinative morphology, and the widespread occurrence of sandhi.In this thesis, a hybrid architecture is proposed that integrates a fine-tuned MuRIL transformer with a character-level convolutional neural network (CNN) for suffix feature extraction. Four distinct fusion strategies – concate-nation, gated fusion, multi-head cross-attention, and low-rank bilinear pooling – are systematically investigated for combining contextual transformer representations with explicit morphological suffix features. A comprehensive ablation study comprising 23 model configurations (7 baselines and 16 hybrid variants across 4 suffix lengths and 4 fu-sion methods) is conducted, with each configuration evaluated over 3 random seeds for a total of 69 experimental runs on the Universal Dependencies Sanskrit–Vedic treebank. The best hybrid configuration (CNN-SL10-attention) achieves a macro-F1 score of 97.47±0.01%, outperforming all baselines including standalone MuRIL (97.27±0.09%), XLM-R (97.39±0.05%), mBERT (97.25±0.01%), and ByT5 (95.70±0.12%). The hy-brid model also demonstrates the lowest cross-seed variance (σ = 0.009) among all evaluated models, representing approximately a tenfold improvement in stability over the MuRIL baseline. The ablation study reveals that attention-based fusion with longer suffix windows consistently yields the best performance. However, the narrow macro-F1 range across all 16 hybrid variants (97.34–97.47%) together with a non-significant McNemar test result between the hybrid model and MuRIL (p = 0.576) suggests that pre-trained multilingual transformers may already encode much of the morphological information captured by explicit suffix features. This finding has broader implications for the design of

Browse

Browsing Department of Computer Science and Engineering by Title

Results Per Page

Sort Options