BITS Faculty Publications

Permanent URI for this communityhttp://localhost:4000/handle/123456789/1867

Browse

Subject: search.filters.subject.Deep learning

Search Results

Now showing 1 - 10 of 28
  • No Thumbnail Available
    Item
    A quality-of-service-centric uplink rate-splitting approach for next-generation multiple access
    (IEEE, 2025-06) Tripathi, Sharda
    Recently, Rate-Splitting Multiple Access (RSMA) has emerged as a powerful paradigm for meeting the demanding performance requirements of 6G wireless networks through non-orthogonal high-rate data transmission. However, uplink access in RSMA necessitates optimizing the decoding order, which can lead to significant search latency. Besides, the process overlooks the Quality-of-Service (QoS) constraints of different traffic types, making current RSMA methods inadequate, especially for low-latency communication. Here, we address this issue by proposing QORA, short for QoS-aware One-shot Rate-splitting multiple Access, a multi-agent Deep Q-Network (DQN) framework that leverages a novel QoS-aware transmit power allocation and decoding order policy in uplink RSMA that achieves remarkable performance improvements while maintaining low latency and high admission rates.
  • No Thumbnail Available
    Item
    Medical image segmentation using advanced UNETt: VMSE-Unet and VM-Unet CBAM+
    (2025-07) Chalapathi, G.S.S.
    In this paper, we present the VMSE U-Net and VM-Unet CBAM+ model, two cutting-edge deep learning architectures designed to enhance medical image segmentation. Our approach integrates Squeeze-and-Excitation (SE) and Convolutional Block Attention Module (CBAM) techniques into the traditional VM U-Net framework, significantly improving segmentation accuracy, feature localization, and computational efficiency. Both models show superior performance compared to the baseline VM-Unet across multiple datasets. Notably, VMSEUnet achieves the highest accuracy, IoU, precision, and recall while maintaining low loss values. It also exhibits exceptional computational efficiency with faster inference times and lower memory usage on both GPU and CPU. Overall, the study suggests that the enhanced architecture VMSE-Unet is a valuable tool for medical image analysis. These findings highlight its potential for real-world clinical applications, emphasizing the importance of further research to optimize accuracy, robustness, and computational efficiency.
  • No Thumbnail Available
    Item
    A review on WSN based resource constrained smart IoT systems
    (Springer, 2025) Haribabu, K.
    In Wireless Sensor Network (i.e. WSN) based resource constrained Internet of Things (i.e. IoT) environments, efficient data forwarding is achieved through cluster based mechanisms, where cluster heads facilitate communication among themselves and with the sink node. Data collected by each cluster head is temporarily buffered before being transmitted to the sink via multi-hop communication. The integration of advanced wireless technologies, such as 5th Generation (i.e. 5G) networks, offers significant benefits, including reduced latency, extensive coverage, improved spectral efficiency, and higher data transmission rates. Incorporating Device-to-Device (i.e. D2D) communication further enhances energy efficiency and offloads data traffic, addressing critical IoT requirements such as low latency, increased network capacity, and improved spectral and energy efficiency. Software Defined Networking (i.e. SDN) addresses diverse IoT network needs across domains like smart grids, healthcare, traffic signaling, agriculture, and smart homes by enabling efficient communication, network management, and innovative control procedures. However, SDN’s application for anomaly detection and primary defense against security threats in IoT systems remains underexplored. This research investigates the potential of the design of an intelligent mechanism for energy efficient, privacy preserving, and secure communication in WSN based resource constrained IoT systems. The proposed approach leverages advanced technologies such as SDN, Machine Learning (i.e. ML), Deep Learning (i.e. DL), D2D communication, Computer Vision, and Network Function Virtualization (i.e. NFV). Additionally, it emphasizes assessing and offloading specific IoT application functions onto the network’s edge to enhance performance. Moreover, the development of lightweight security mechanisms for secure communication in resource constrained IoT environments is also identified as a crucial research domain.
  • No Thumbnail Available
    Item
    Women sport actions dataset for visual classification using small-scale training data
    (Sage, 2025-07) Bera, Asish
    Sports action classification representing complex body postures and player-object interactions, is an emerging area in image-based sports analysis. Some works have contributed to automated sports action recognition using machine learning techniques over the past decades. However, sufficient image datasets representing women’s sports actions with enough intra- and inter-class variations are not available to the researchers. To overcome this limitation, this work presents a new dataset named WomenSports for women’s sports classification using small-scale training data. This dataset includes a variety of sports activities, covering wide variations in movements, environments, and interactions among players. In addition, this study proposes a convolutional neural network (CNN) for deep feature extraction. A channel attention scheme upon local contextual regions is applied to refine and enhance feature representation. The experiments are carried out on three different sports datasets and one dance dataset for generalizing the proposed algorithm, and the performances on these datasets are noteworthy. The deep learning method achieves 89.15% top-1 classification accuracy using ResNet-50 on the proposed WomenSports dataset, which is publicly available for research at Mendeley Data.
  • No Thumbnail Available
    Item
    Bayesian deep learning meets self-attention: a risk-aware approach to advertisement optimization
    (IEEE, 2025-05) Bhatia, Ashutosh; Tiwari, Kamlesh
    In the highly competitive landscape of e-commerce advertising, maximizing Return on Advertising Spend (ROAS) is critical, yet remains inherently uncertain due to auction-based bidding dynamics and fluctuating market conditions. Traditional deterministic models fail to capture this uncertainty, necessitating a probabilistic approach that balances predictive accuracy with interpretability. To address this challenge, the paper proposes a novel Hierarchical Bayesian Deep Learning framework that integrates a Bayesian Belief Network (BBN) for structured probabilistic reasoning and a Mixture Density Network (MDN) for full distributional modeling of ROAS. The BBN models dependencies among campaign variables, offering interpretable insights, while the hierarchical deep learning architecture overcomes scalability limitations in high-dimensional settings through self-attention mechanisms. Experiments demonstrate up to 22.8% lower RMSE and 27.4% better Negative Log Likelihood (NLL) and up to 31.2% lower Kullback-Leibler divergence (KLD) than state-of-the-art methods (DeepAR, Prophet, NGBoost), achieving an R2 of 98% with an inference speed of 5.2 ms per campaign, making real-time bidding feasible. Ablation studies confirm that attention-driven feature selection and calibrated uncertainty quantification significantly enhance both predictive performance and explainability, identifying key drivers of campaign success. By providing precise, uncertainty-aware, and explainable predictions, this approach enables adaptive bidding strategies, optimized budget allocation, and risk management, setting a new benchmark for intelligent decision-making in digital advertising.
  • No Thumbnail Available
    Item
    Deep learning approaches for driver distraction detection using driver facing cameras: literature review and empirical study using cnn classifiers on a 100-driver image dataset
    (2025-05) Bhatia, Ashutosh; Sharma, Yashvardhan; Tiwari, Kamlesh
    Distracted driving contributes to thousands of fatalities and injuries globally. According to India’s Ministry of Road Transport and Highways (MoRTH), distraction-related behaviors such as rear-end and off-road collisions accounted for nearly one-fourth of all traffic incidents in 2022. The U.S. National Highway Traffic Safety Administration (NHTSA) reported 3,275 deaths and over 324,000 injuries from distraction-related crashes in 2023. In Europe, the European Road Safety Observatory (ERSO) observed handheld phone use by drivers in up to 9.4% of vehicles across member states, with self-reported texting rates reaching 53%. Despite numerous studies and surveys on driver distraction detection, existing literature remains fragmented, often combining multiple sensor modalities or distraction with related driver states such as fatigue. Prior empirical efforts also lack a unified benchmarking strategy to assess model generalization under shifts in viewpoint or spectral input. This paper presents a focused survey and empirical study of visiononly distraction detection using deep learning models applied to driver-facing camera inputs. It introduces a conceptual model linking behavioral cues to cognitive distraction, defines the visionbased Driver Distraction Detection (vDDD) system with alert logic, and develops structured taxonomies of datasets, architectures, and learning strategies. Using the 100-Driver dataset, the empirical study evaluates 26 CNN classifiers under 64 crossdomain configurations, systematically analyzing generalization across modality and camera view changes. Results show that frontal RGB-trained models generalize better than their NIRtrained counterparts and that lightweight models trade off accuracy under rare class scenarios for faster inference. The study establishes the vDDD paradigm as a vision-based behavioral modeling approach for distraction detection using driver-facing camera data. It outlines future research directions in spectrumaligned augmentation, attention modeling, and lightweight visuallanguage fusion, emphasizing deployment-focused strategies such as quantization, contrastive learning, and progressive fine-tuning.
  • No Thumbnail Available
    Item
    Efficient edge AI implementation for IoT device identification for hierarchical federated learning
    (Inder Science, 2025-03) Shenoy, Meetha V.
    As IoT devices proliferate, efficient IoT device identification is crucial for resource management, planning, and detecting anomalous traffic. Traditional ML-based identification relies on centralised training, but federated learning (FL) offers a privacy-preserving alternative, enabling collaborative model training without sharing raw data. FL enhances edge devices' ability to identify previously unconnected devices. However, resource constraints like limited computation, power, and communication capabilities may prevent some edge devices from actively participating in FL. We propose a solution where resource-limited IoT devices benefit from FL by subscribing to server-based services. This work presents an efficient AI model implementation for IoT device identification on embedded edge devices, detailing the toolflow from model generation to hardware implementation. We apply and evaluate various model optimisation techniques to balance performance and resource trade-offs, offering insights to advance edge-AI and scalable FL-based ML applications for IoT networks.
  • No Thumbnail Available
    Item
    A detailed comparative analysis of automatic neural metrics for machine translation: bleurt & bertscore
    (IEEE, 2025-04) Chamola, Vinay; Gupta, Karunesh Kumar
    Bleurt a recently introduced metric that employs Bert, a potent pre-trained language model to assess how well candidate translations compare to a reference translation in the context of machine translation outputs. While traditional metrics like Bleu rely on lexical similarities, Bleurt leverages Bert's semantic and syntactic capabilities to provide more robust evaluation through complex text representations. However, studies have shown that Bert, despite its impressive performance in natural language processing tasks can sometimes deviate from human judgment, particularly in specific syntactic and semantic scenarios. Through systematic experimental analysis at the word level, including categorization of errors such as lexical mismatches, untranslated terms, and structural inconsistencies, we investigate how Bleurt handles various translation challenges. Our study addresses three central questions: What are the strengths and weaknesses of Bleurt, how do they align with Bert's known limitations, and how does it compare with the similar automatic neural metric for machine translation, BERTScore? Using manually annotated datasets that emphasize different error types and linguistic phenomena, we find that Bleurt excels at identifying nuanced differences between sentences with high overlap, an area where BERTScore shows limitations. Our systematic experiments, provide insights for their effective application in machine translation evaluation.
  • No Thumbnail Available
    Item
    Optimizing liquid neural networks: a comparative study of ltcs and cfcs
    (IEEE, 2024) Challa, Jagat Sesh
    Liquid Time Constant Networks (LTCs) and Closed Form Continuous Networks (CFCs) are recent time-continuous RNN models known for superior expressivity and efficiency in time-series prediction and autonomous navigation. This paper provides an accessible overview of these models and investigates their performance on tasks like Atari ’Breakout’ behavior cloning, steering angle prediction, and Global Horizontal Irradiance (GHI) forecasting. We optimize LTC and CFC cells within network structures, comparing them with LSTM. Detailed experiments highlight the impact of various hyperparameters, underscoring the effectiveness of LTCs and CFCs in dynamic prediction tasks.
  • No Thumbnail Available
    Item
    Deep learning based techniques to develop & enhance assistive gear for visually impaired
    (IEEE, 2024-12) Rout, Bijay Kumar
    This work investigates existing solutions tailored for the visually impaired, focusing on economically viable options for non-first-world communities. The exploration involves developing a real-time obstacle-tracking model using the YOLO (You Only Look Once) algorithm and Text-to-Speech synthesis to provide auditory cues. This effort yields improvements in assistive technology, though it still faces limitations in algorithmic precision and user feedback integration. The research paves the way for refining this technology and envisions its seamless integration into the daily lives of the visually impaired. The findings enhance the performance of assistive technologies, especially for distances less than 1.5 meters. The results show an inaccuracy of less than 10%, translating to a margin of 10−15 cm for objects located one meter away. This work thus provides increased independence and confidence for individuals with visual impairments in navigating and interacting with their surroundings.