Department of Electrical and Electronics Engineering

Permanent URI for this collectionhttp://localhost:4000/handle/123456789/1925

Browse

Has files: NoSubject: search.filters.subject.Neural networks

Search Results

Now showing 1 - 10 of 16
  • No Thumbnail Available
    Item
    A multi-modal smart switching based image transmission using semantic communication
    (IEEE, 2025-02) Tripathi, Sharda; Joshi, Sandeep
    The conventional paradigm of communication primarily concentrates on the transmission of raw data, often disregarding its contextual meaning. However, to tackle the exponential growth in data demands along with the limited availability of transmission bandwidth, there is an increasing need to transition from Shannon’s classical information-theoretic communication to a more advanced framework centered on semantics. This work presents a multi-modal semantic-based communication method for the transmission of high-definition images aimed at optimizing the transmitted data volume while maintaining a high throughput and mean intersection over union score. To this end, two architectural models are explored: a denser ResNet-based and a lightweight U-Net-based. Depending on the required QoS and resource availability, the raw image is either semantically segmented to obtain a fine-grained, pixel-level classification of the image or represented as label semantics, which provides only a higher-level, object-based, or region-based classification prior to its transmission. The experimental results show that such an adaptive semantic image processing approach leads to around 63% reduction in the transmitted data volume without compromising on the quality of image reconstruction.
  • No Thumbnail Available
    Item
    Lightweight convolutional neural network architecture implementation using TensorFlow lite
    (Springer, 2023-06) Asati, Abhijit
    Recently, with the increase in the precision of convolutional neural networks (CNN) on a wide variety of classification and recognition tasks, the demand for their deployment has dramatically increased. Even the focus is on lightweight, faster, and low-power implementations. In this paper, we have implemented a CNN model onto an embedded platform, ‘Raspberry Pi 4-Model B edge computing system (RP4-BECS)’. This CNN model was initially trained and verified in MATLAB and then implemented on the Machine Learning (ML) framework to generate a TensorFlow lite (TF-lite) flat buffer format. This implementation offers a reduced size of models with good prediction accuracy and lesser inference time as compared with the available literature. We attempted three trials for all the digits from 0 to 9 to evaluate average prediction accuracy and average inference time. An average prediction accuracy of 99.32% and average inference time of 22.53 ms is achieved for the Sign Language Digits Database (SLDD). Further, an average prediction accuracy of 99.09% and average inference time of 13.28 ms is achieved for the Modified National Institute of Standards and Technology Database (MNIST). The model sizes implemented using TF-Lite are highly reduced to 1.53 MB for SLDD and 148 KB for the MNIST database. The obtained accuracy, inference time and model sizes are better than published results.
  • No Thumbnail Available
    Item
    Sensor Information Processing for Wearable IoT Devices
    (Springer, 2019-11) Shenoy, Meetha V.
    Sensing technology is one of the core enablers of IoT and the improvement in sensing technology has lead to the proliferation of small form-factor, cost-effective and accurate sensors for wide variety of wearable applications. With wearable devices receiving widespread acceptance, their requirements are becoming more demanding, with the focus shifting from simple monitoring to context aware intelligent devices. This chapter presents a comprehensive description of the technical opportunities and challenges in the design of sensor information processing systems for wearables. A systematic survey of the state of the art architectures for sensor fusion for different application classes of wearable’s is presented. A discussion on design considerations for architecting sensor processing systems, including hardware, networking protocols, and algorithms at the edge, cloud level is provided. The chapter is concluded with a discussion on innovation directions in smart sensing and information processing in wearable devices.
  • No Thumbnail Available
    Item
    Indoor localization in NLOS conditions using asynchronous WSN and neural network
    (IEEE, 2017) Shenoy, Meetha V.
    Development of technologies for accurate localization of objects in indoor environments can transform wide application domains like healthcare, warehouses and fitness industries. In this paper, we present a novel neural network and asynchronous wireless sensor network (WSN) based indoor localization scheme. A custom designed ultrasonic trans-receiver serves as the back bone of the localization scheme. In addition to the experiments on hardware, we utilize Locusim, an acoustic simulator to augument extensive analysis on Non line of Sight (NLOS) conditions. We demonstrate that the neural network based localization scheme can provide an accuracy suitable for most of the real world applications even under NLOS conditions.
  • No Thumbnail Available
    Item
    Automatic Control of an Asymmetric Fighter Aircraft Performing Supermanoeuvres
    (AIMT, 2020) Mukherjee, Bijoy K.
    Center-of-gravity (c.g.) of a combat aircraft may deviate significantly from the plane of symmetry due to asymmetric release of stores, leading to a highly coupled asymmetric six degree-of-freedom (6-DOF) dynamics. Additional nonlinearity and cross-coupling between the longitudinal and lateral-directional dynamics result when the aircraft attempts some supermanoeuvres under such asymmetric conditions. This renders nonlinear control implementation almost unavoidable for the safety of the aircraft. However, success of such control schemes heavily depends on the accurate onboard information of the actual asymmetric c.g. location. The present paper proposes a novel neural network aided sliding mode based hybrid control scheme which does not require such online c.g. information at all. The neural controller part is trained offline so that it can compensate for the deviations in the aircraft dynamics arising from the lateral mass asymmetry and the sliding controller is designed assuming the nominal or symmetrical dynamics to execute the intended manoeuvres. To validate the usefulness of the proposed control scheme, two well-known supermanoeuvres cobra and Herbst are simulated and it is shown that the manoeuvre performance does not get affected appreciably even under a wide range of lateral c.g. movements.
  • No Thumbnail Available
    Item
    Understanding the Use and Abuse of Social Media: Generalized Fake News Detection With a Multichannel Deep Neural Network
    (IEEE, 2022-11) Chamola, Vinay
    Fake news has spread across social media platforms and with the ease of access, negative consequences have come with it on individuals and society. This issue has become a focus of interest among various research communities, including artificial intelligence (AI) researchers. Existing AI-based fake news detection techniques primarily make use of a 1-D convolutional neural network (1D-CNN) with unidirectional word embedding. We propose a multichannel deep convolutional neural network (CNN) with different kernel sizes and filters as an AI technique. Multiple embedding of the same dimension with different kernel sizes technically allows the news article to be processed at different resolutions of different n-grams at the same time. Different kernel sizes increase the learning ability of the proposed classification model. The proposed model determines how to integrate these interpretations (different n-grams) most suitably. Three real-world fake news datasets were used in experiments to validate the classification performance. The classification results showed that the proposed model has high accuracy in detecting fake news. Regardless of the dataset, the proposed model can be used for fake news detection in binary classification problems
  • No Thumbnail Available
    Item
    Traffic Jam Probability Estimation Based on Blockchain and Deep Neural Networks
    (IEEE, 2021-07) Chamola, Vinay
    The exponential surge in the number of vehicles on the road has aggravated the traffic congestion problem across the globe. Several attempts have been made over the years to predict the traffic scenario accurately and consequently avoiding further congestion. Crowdsourcing has come forward as one of the most adopted methods for predicting traffic intensity using live data. However, the privacy concerns and the lack of motivation for the live users to help in the traffic prediction process have rendered existing crowdsourcing models inefficient. Towards this end, we present an advanced blockchain-based secure crowdsourcing model. Not only does our model ensure privacy preservation of the users, but by incorporating a revenue model, it also provides them with an incentive to participate in the traffic prediction process willingly. For accurate and efficient traffic jam probability estimation, our work proposes a neural network-based smart contract to be deployed onto the blockchain network. The results reveal that the proposed model is highly efficient in terms of attaining high participation and consequently obtaining highly accurate predictions.
  • No Thumbnail Available
    Item
    Audio classification using braided convolutional neural networks
    (IET, 2020-09) Ajmera, Pawan K.
    Convolutional neural networks (CNNs) work surprisingly well and have helped drastically enhance the state-of-the-art techniques in the domain of image classification. The unprecedented success motivated the application of CNNs to the domain of auditory data. Recent publications suggest hidden Markov models and deep neural networks for audio classification. This study aims to achieve audio classification by representing audio as spectrogram images and then use a CNN-based architecture for classification. This study presents an innovative strategy for a CNN-based neural architecture that learns a sparse representation imitating the receptive neurons in the primary auditory cortex in mammals. The feasibility of the proposed CNN-based neural architecture is assessed for audio classification tasks on standard benchmark datasets such as Google Speech Commands datasets (GSCv1 and GSCv2) and the UrbanSound8K dataset (US8K). The proposed CNN architecture, referred to as braided convolutional neural network, achieves 97.15, 95 and 91.9% average recognition accuracy on GSCv1, GSCv2 and US8 K datasets, respectively, outperforming other deep learning architectures.
  • No Thumbnail Available
    Item
    Analysis of Adversarial Jamming From a Quantum Game Theoretic Perspective
    (IEEE, 2023-03) Bitragunta, Sainath
    Over recent years, quantum communication systems have demarcated themselves as promising candidates for deployment in next-generation communication networks (6G and beyond). Several recent experimental demonstrations of such complex systems have been highly successful and have been instrumental in transitioning this field from the theoretical to the practical domain. In this article, we investigate the application of quantum game theory for the modeling and analysis of jamming in the context of quantum networks. We begin with a general model of jamming based on the Colonel Blotto game and generalize it to the context of quantum networks. We provide an in-depth analysis of the two-person quantum Colonel Blotto game (QCBG) in relation to classical versus classical, classical versus quantum, and quantum versus quantum strategies. We also investigate the Nash equilibria for such games via a multiagent adversarial reinforcement learning-based system. Finally, we discuss further optimizations on this model and outline several open problems for further research along these lines.
  • No Thumbnail Available
    Item
    Performance enhancement of neural network training using hybrid data division technique for photovoltaic power prediction
    (IEEE, 2017) Kumar, Rajneesh
    The data available for training, testing and validation of a neural network defines the efficiency or performance of the network. This research work compares the data division techniques like random division, Self-Organizing Maps, fuzzy c means and K-means to predict power output of a solar panel under loss conditions. The data used is obtained from a series of experiments on a soiled panel. Finally, a new data division technique for designing neural networks in PV module output prediction is proposed and its efficiency is compared with other discussed data division techniques. The proposed data division technique helps in building a better neural network model with comparatively less data available.