National Institute of Technology, Patna

Underwater image enhancement using dehazing approach represents significant research challenges in the field of underwater imaging and computer vision due to the effect of light absorption, scattering and wavelength dependent attenuation. This result in reduced contrast, color distortion, and blurred image, thus making it harder to observe, monitor and navigate underwater imaging. This research provides a comprehensive analysis of underwater image dehazing, describing the evolution of underwater image dehazing from traditional physics-based and enhancement-based techniques to advanced deep learning and transformer-based techniques. The work reviews the fundamental optical principles, such as Beer–Lambert equation and the Jaffe–McGlamery imaging model, which explain the concept of light wave propagation and its attenuation in underwater environments. Traditional enhancement-based models, such as histogram equalization, contrast-limited adaptive histogram equalization, retinex-based, and multi-scale fusion techniques, are efficient and have lower computational complexity. Recent deep learning-based models, comprising convolutional neural networks, generative adversarial networks, and transformer-based architectures, restore fine details, enhance perceptual quality, and improve generalisation under varied environmental conditions. This comprehensive review also highlights the performance assessment of existing benchmark datasets and image quality measurement parameters such as peak signal-to-noise ratio, structural similarity index, underwater image quality measure, and underwater color image quality evaluation. It also states various underwater image enhancement challenges such as changing environmental effect, data limitations and restrictions on real-time execution. It additionally glances at real time applications in robotics, underwater archaeology and marine biology which are based on emerging trends like physics-guided learning,, self-supervised training, lightweight architectures and domain adaptation.

With the rapid growth of digital content on the web, search engines and recommendation systems (RS) have become important tools to search and discover meaningful information efficiently. Traditional RSs often struggle to generate accurate and personalized recommendations due to the overwhelming expansion of online data, leading to information overload. Therefore, a better understanding of modern deep learning-based solutions is required to overcome these limitations. Deep learning techniques have emerged as a powerful method, utilizing their ability to model complex, nonlinear relationships and extract meaningful patterns from multidimensional datasets. Recent advancements such as Graph Neural Networks (GNNs), Transformers, and Reinforcement learning have further improved recommendation accuracy by capturing intricate user–item interactions and long-range dependencies. Consequently, deep learning-based RSs have gained popularity due to their ability of modeling user preferences, behaviors, and item characteristics at multiple levels, enhancing accuracy, personalization, and scalability. The objective of this review is to systematically explore these developments and provide an understanding of current trends. This review systematically studies various state-of-the-art deep learning-based RSs, focusing on their architectures, implementation progress, and practical applications. We analyze their effectiveness in extracting intrinsic user and item features, examine their strengths and limitations, and highlight emerging trends in the field. The primary contribution is to compare traditional and deep learning-based recommendation approaches with their key challenges and future research opportunities. Additionally, we discuss fundamental issues related to traditional and deep learning-based RS, which continue to hinder the widespread adoption of deep learning techniques in RSs. Overall, this survey presents a structured framework to guide future research and support the development of more robust, scalable, and efficient deep learning methodologies for RSs.

The COVID-19 pandemic has significantly altered global health, economies, and societal structures, making accurate daily case predictions essential for strategic planning and outbreak control. This study employs two machine learning models: a multi-layer perceptron (MLP) neural network and a bio-inspired sperm swarm optimization (SSO) algorithm hybridized with MLP (MLP-SSO), comparing their performance in forecasting daily COVID-19 deaths in Brazil, India, Russia, and the United States using time series data from January 20 to September 15, 2020. To ensure a realistic evaluation of the model’s forecasting capability, the data were chronologically split into training (first 70

The rapid expansion of mobile connectivity and the global reliance on smartphones have positioned Android as the leading platform, driven by its affordability and open source framework. However, its open architecture also introduces vulnerabilities, making it a prime target for malware and posing severe cybersecurity risks. To address these issues, this research introduces an improved, static-analysis-based malware detection approach based on reverse-engineered, disassembled code (smali files), focusing on real permissions and Application Programming Interface (API) call sequence features. The proposed methodology utilizes the Bidirectional Encoder Representations from Transformers (BERT) model to convert textual features into contextualized vector representations, effectively capturing semantic relationships. Particle Swarm Optimization (PSO) with nested cross-validation reduces dimensionality by 50

A phase change material-based system that can absorb and release energy is a crucial component in renewable energy applications. The performance of such a system can be improved by incorporating fins into the phase change material region of the system. Conventionally, the most common shape for these fins is straight and unbend. However, in this study, the impact of using wave-shaped bend fins has been explored instead of the unbend ones. Different cases have been analysed, with variations in their amplitude, which influences the degree of bendiness. For the considered cases, a transient analysis computer simulation model was developed and validated against published experimental data. Using this model, the charging and discharging processes of the PCM were analysed based on liquefied portion and temperature variation. The findings demonstrate that bend fins significantly enhance heat transfer and PCM flow dynamics. For the wave-shaped fin with the highest amplitude, the charging and discharging times were reduced by 35.55% and 37.82%, respectively, compared to straight unbend fins. The thermal-fluid behaviour of the material was further evaluated based on phase change time, as well as variations in temperature and velocity of the PCM at critical points (locations below the first and second peaks of the horizontally oriented wave-shaped bend fin). The results reveal that increased fin bendiness (higher amplitude of the wave-shaped fin) delays the initiation of phase change at these critical points. Additionally, PCM flow dynamics near critical points were significantly improved with bend fins.