Detection and tracking of small objects like vehicles, aeroplanes and ships from satellite imagery targets several critical applications in urban planning, air and sea traffic management, marine surveillance, disaster prediction and relief, etc. But accurately locating small-objects in large and complex satellite images is highly challenging due to loss of contextual information, blurred objects, diminished sizes, etc. This paper proposes a novel dual attention-integrated YOLO model to accurately detect small-objects in the stand Skyfusion dataset. The selected baseline YOLOv11s is fused with global attention and Efficient channel attention mechanisms at varying feature resolutions for retaining global contextual and spatial information along with channel inter-dependencies. The proposed work applies targeted data augmentation to oversample the Ship class instances and benchmarks the proposed technique against state-of-the-art models with the highest F-score and mAP of 72.4% and 71.3% . With minimal increase in parameters, the proposed model thus serves as a scalable and reliable solution for tiny object detection in challenging aerial imagery.
Biological processes are executed concurrently and on different scales. Modeling of biological processes becomes challenging when the concentrations of biochemicals are imprecise. A tractable formalism for model construction to address imprecision in the concentrations of biochemicals. The model calibration is performed by generating finer models from coarser models that represent the set of sequences of biochemical reactions. Learning of the specifications extracted from the coarser model is applied effectively to construct a finer model. Bounded model checking is performed on the prototype by addressing the uncertainty in the concentration of biochemical reactions. Learning of temporal logic formula is leveraged to address uncertainty in the model. The computational feasibility of the formalism is evaluated on a prototype of the RKIP-inhibited ERK pathway. Results are presented and formalism is promising.
Product discovery is crucial for any retailer. Traditional approaches to achieve this involve understanding queries. Current methods for query understanding require knowledge of product attributes. However, the attributes of interest may not be known beforehand. In addition, these attributes may evolve over time. This study aims to identify significant attributes of the product from descriptions, reviews, and other sources such as articles about the products. A phrase mining technique is used to extract relevant and important words and phrases. Meaningful concepts are generated by clustering these words and phrases. Knowledge of existing product attributes and their corresponding values is utilized as constraints during the clustering process. To uncover new attributes, we utilize topic modeling. This study presents a novel topic modeling framework that explicitly handles constraints. Experimental results from real-world examples and a benchmark dataset demonstrate the efficacy of the framework.
Developing robust conversational agents for e-commerce, particularly for voice-activated product ordering, presents a significant challenge. This paper introduces a lightweight and efficient framework for this task, centered on a compact ( ∼ 1B parameter) Large Language Model (LLM), Gemma-3 1B. The proposed methodology leverages parameter-efficient fine-tuning (PEFT) through Low-Rank Adaptation (LoRA) to specialize the model for semantic parsing, combined with post-training quantization to ensure efficient deployment on both server and mobile platforms. The system is trained on a synthetic dataset of 20,000 spoken shopping requests, enabling it to map diverse user utterances to precise JSON order specifications. Experimental evaluation demonstrates the high efficacy of this approach. The LoRA-tuned model achieves an exact JSON match accuracy of approximately 94.6 ∼ 0.25 s) and feasible on-device latency ( ∼ 2.5 s) when using 8-bit quantization. These findings demonstrate that compact LLMs, when coupled with LoRA-based specialization and quantization, can serve as both accurate and deployable semantic parsers for voice-based commerce applications. The proposed framework not only achieves competitive accuracy with significantly reduced hardware requirements but also establishes a reproducible blueprint for real-world conversational AI deployment.
This study presents a comprehensive computational analysis of a curated dataset of instruction–response pairs within the sustainable fashion domain. The dataset comprises consumer-style queries and expert-style answers, providing a dual perspective view of sustainability discourse in fashion. Using a multi-stage text mining pipeline, we examine linguistic properties, sentiment polarity, sustainability depth, inclusivity attributes, and thematic structures through non-negative matrix factorization (NMF) topic modeling. Sentiment analysis is implemented using a domain-adapted lexicon to capture tone differences between queries and responses, while sustainability depth scoring evaluates the presence of critical concepts such as material choice, care practices, circular economy principles, ethical sourcing, and environmental impact. Inclusivity is assessed through rule-based detection of body-positive, gender-neutral, cultural, and sensory accessibility references. The results reveal a marked positivity bias in responses compared to instructions, with a consistent tone shift towards encouragement and reassurance. Topic modeling identifies ten recurring themes, including materials and fabrics, care and longevity, capsule wardrobe planning, ethics and sourcing, and seasonal dressing. Sustainability depth is highest in material-focused and ethics-related topics, whereas styling and fit-oriented topics show lower integration of environmental principles. Inclusivity coverage is uneven, with cultural and gender-neutral aspects underrepresented. The findings are interpreted in the context of strategic brand communication, consumer education, and policy advocacy.
Accurate animal identification is vital for precision livestock farming, yet existing biometric methods often struggle with subtle inter-animal similarities and the lack of public datasets, especially for indigenous breeds. This study introduces a curated bovine facial dataset comprising 3,680 images from 184 Indian cows, generated through automated video frame extraction, sharpness-based filtering, YOLOv11 face detection, and clustering for quality and diversity. Three CNN architectures—ResNet-50, Inception-V3, and EfficientNet-B4—were evaluated using cross-entropy loss. EfficientNet-B4 achieved the highest closed-set accuracy (97.28
This study presents a YOLOv11-based framework for automated assessment of bread freshness over a period of ten days. A comprehensive dataset of 3,600 bread samples collected from 20 bakeries was created to capture preservative effects and spoilage progression, enabling precise edibility grading. An advanced preprocessing pipeline employing YOLOv11-based region-of-interest (ROI) extraction ensured analysis focused on relevant bread portions. The proposed framework achieved strong segmentation and classification performance, with mAP50 scores of 0.965 for both bounding box and mask predictions and an overall classification accuracy of 81
Software complexity has been long recognized as a key determinant of quality in large scale software development. However, complexity is multi-faceted and understanding how it relates to the number of defects in a software system has its challenges. Empirical studies on software complexity are mostly correlational in nature. In this paper, we examine two different aspects of complexity from a causal perspective and examine their relation to defects using a causal discovery based approach. The discovered causal model is validated using a large-scale development. Validation results, along with software engineering domain knowledge yield notable insights that are relevant for the design, development, and maintenance of robust and resilient software systems. To the best of our knowledge, our approach is among the pioneering studies that apply causal discovery techniques in understanding how code level complexity influences defect occurrence in real-world software development.
Intracranial Hemorrhage (ICH) represents a life-threatening neurological condition that demands immediate and precise diagnosis for timely clinical intervention. However, most existing automated detection systems struggle with issues such as unbalanced datasets, poor adaptability across data sources, and insufficient attention to hemorrhage-specific brain regions. To overcome these limitations, this study introduces a comprehensive dataset for ICH detection that integrates newly acquired hospital images with existing open-access datasets, encompassing six major subtypes of hemorrhage. A novel two-stage ensemble framework is proposed: in the first stage, image quality is enhanced using Contrast Limited Adaptive Histogram Equalization (CLAHE); in the second stage, a Brain Region Detector based on You Only Look Once version 11 (YOLOv11) is employed to accurately localize potential hemorrhagic regions. From these localized areas, Histogram of Oriented Gradients (HOG) and Local Binary Pattern (LBP) features are extracted, fused, and normalized through Min–Max scaling. The feature space is then optimized using Linear Discriminant Analysis (LDA) to enhance class separability. Finally, a Support Vector Machine (SVM) classifier with a Radial Basis Function (RBF) kernel is used for classification, achieving a 96.13
Electroencephalogram (EEG) signals are widely employed in stress recognition research due to their ability to capture subtle neural variations associated with mental states. However, EEG signals are often corrupted by various artifacts and noise sources, including ocular, muscular, and environmental interferences, making preprocessing a critical step. This paper presents a comparative evaluation of classical preprocessing approaches (such as filtering and conventional wavelet denoising) against the Adaptive Wavelet Transform (AWT) framework for EEG denoising. The study investigates the effectiveness of these methods based on Root Mean Square Error (RMSE), and preservation of signal energy, focusing on their role in enhancing the discriminability of stress-related EEG patterns. Experimental results on benchmark EEG datasets demonstrate that AWT provides the average MSE value of 0.75e−07, as compared to 2.9e−5 and 3.55e−6 for band-pass filter and Wavelet transform, respectively. Similarly, average retained energy in de-noised signal using AWT is 96.66
We present a deterministic parallel external-memory (PEM) algorithm that sorts N integers from {1,2,… ,l} . The algorithm sorts N integers from [1, l] in O( log l/log m( M/B + log _M/BN/B) ) parallel steps and O( N/B·log l/log (M/B)) I/O operations. on the Concurrent Read Exclusive Write PEM model. (Here M is the size of the internal main memory of processors and B is the block size for I/Os.)
Satisfiability (SAT) is a fundamental problem in computer science. This paper investigates a specific subclass of satisfiable expressions: those with a unique satisfying assignment (USAT). We introduce the concept of minimal USAT patterns — graph structures that represent USAT expressions which cease to be USAT if any clause is removed. We outline a method for identifying and cataloguing these patterns, present empirical findings from a database of them, and examine their practical applications. Finding these minimal patterns in larger SAT instances can serve multiple purposes, such as a preprocessing step to simplify the problem, a generalisation of a unit clause during the execution of a SAT solver, or as a distinctive feature in classifying SAT instances.
Detecting and neutralizing tanks from long distances using Unmanned Aerial Vehicles (UAVs) is an emerging research area with the potential to significantly enhance military capabilities and reduce loss of life. Despite this potential, the domain remains relatively underexplored, especially in specific terrains and operational contexts. This study presents the development of a real-time tank detection and targeting system. The proposed end-to-end framework integrates advanced machine learning algorithms, image processing techniques, and precision targeting mechanisms. The system operates in three phases: Phase 1 detects tanks using object detection algorithms such as YOLOv7; Phase 2 enhances detection accuracy at long distances through the Slicing-Aided Hyper Inference (SAHI) algorithm; and Phase 3 applies the StrongSORT tracking algorithm for continuous target tracking and engagement. Experimental evaluation on the VisDrone dataset demonstrates the system’s effectiveness in reducing operational risks and improving mission success rates, achieving a mean Average Precision (mAP) of 0.88 at IoU 0.5, with performance decreasing to 0.59 for mAP@0.5:0.95.
Synchronized arrival coordination represents a critical challenge in multi-agent systems, particularly for military tactical operations, emergency response, and autonomous swarm robotics, where precise temporal coordination is essential. Traditional Ant Colony Optimization (ACO) algorithms optimize for shortest paths but fail to address the temporal synchronization requirements inherent in multi-agent coordination scenarios. We introduce MACO-Sync (Multi-Agent Ant Colony Optimization for Synchronized Arrival), a novel ACO variant that explicitly optimizes for temporal coordination rather than path minimization. Our key contribution is a synchronized pheromone update mechanism based on arrival time patterns. The agents’ pheromone contributions are weighted by their synchronization variance with other agents. Extensive experimental evaluation on tactical coordination scenarios demonstrates that MACO-Sync achieves 23.9 times better synchronization scores and 6.8 times lower arrival variance compared to baseline algorithms. Our approach enables effective multi-agent coordination without explicit inter-agent communication. This is particularly suitable for military and emergency response applications where communication constraints are prevalent. Code: GitHub repository .
Multi-head self attention provides strong representational performance but is computationally expensive, which limits its applicability in low-resource environments. Although Multi-Head Latent Attention alleviates the key-value cache bottleneck by compressing representations into a latent space, model training remains computationally demanding, particularly under resource constraints. Our objective is to further reduce the number of learnable parameters in Multi-Head Latent Attention by pre-multiplying the query and key weights, since these matrices are ultimately multiplied during attention computation and by introducing a modified architecture that incorporates a latent space for both the queries and key-value latent representation weights. We refer to this approach as Multi-Head Multi-Latent Attention (MMLA). The experimental results indicate a performance comparable to MSA, accompanied by a 47
The proliferation of Internet of Things (IoT) networks has increased exposure to sophisticated cyberattacks, demanding robust intrusion detection systems (IDS). This paper presents a unified, cross-model performance analysis of five deep learning (DL) frameworks—Transformer-based IDS, Graph Neural Networks (GNN), Conditional Variational Autoencoder (CTVAE), SimCLR-based contrastive learning, and Federated MLP—under a common preprocessing and evaluation protocol. Experiments on two recent real-world datasets (RT-IoT 2022 and ACI IoT 2023) are used to evaluate the performance matrices. Transformer and SimCLR models achieve up to 99
Urban traffic congestion remains a critical challenge, largely due to static signal control systems that fail to adapt to real-time traffic dynamics. We propose a multi-agent traffic signal control framework that integrates Graph Neural Networks (GNNs) with the Soft Actor-Critic (SAC) algorithm under a Centralized Training with Decentralized Execution (CTDE) paradigm. Each intersection is modeled as an agent that learns to optimize signal phases based on local observations enriched with neighborhood context through a graph attention encoder. The reward function balances local intersection efficiency with global throughput and prioritizes public transport. We evaluate our approach in the Bologna-Pasubio scenario using the SUMO simulator. Our method shows its potential for adaptive and coordinated traffic control since it reduces average travel time by 87.3
In industrial environments, robot navigation requires collision-free path planning, especially in the presence of dynamic obstacles such as moving machines, humans, and other autonomous systems. Traditional algorithms like A* and Analytic Hierarchy Process (AHP) are widely used, to solve these problems but they have their limitations. A* primarily focuses on static, distance-optimized path planning but is unable to adapt to dynamic changes. On the contrary, AHP alone is unable handle complex path planning scenarios. In this work, we propose a hybrid algorithm that integrates AHP with Proximal Policy Optimization (PPO), to overcome these limitations. PPO is used for global path planning. By using optimal navigation policies through interaction with the environment, PPO refines the way-points generated by AHP based on distance, angle of movement, and collision safety. The outcome of the hybrid algorithm is an efficient and adaptive set of global way-points which forms the final path. The performance of the proposed AHP-PPO algorithm is compared with the existing methods, such as A* with AHP (AAHP) and Artificial Potential Field (APF), using matrices like path length, angular variation, and obstacle avoidance. AHP-PPO is observed to outperform the existing approaches while demonstrating enhanced adaptability, smoother trajectories, and greater safety.
The immense growth in e-Commerce has intensified the need for efficient last-mile delivery systems. This paper introduces an enhanced Two-Echelon Vehicle Routing Problem with Drones (2E-VRPD) that addresses critical limitations by hyperparameter tuning using Bayesian optimization. A new algorithm is proposed that extends the original Drone Truck Route Construction (DTRC) algorithm by incorporating clustering for customer segmentation and Bayesian optimization using the OPTUNA framework for dynamic synchronization of trucks and drones. Experiments with benchmark instances and real-world case studies demonstrate an average reduction of 9.6
Smart electric grids produce massive volumes of time-series data that must be monitored continuously to ensure operational reliability. This paper proposes an adaptive anomaly detection framework using Gaussian Mixture Models (GMM) to identify abnormal power consumption patterns in individual household grids. By leveraging temporal features and statistical modeling, our approach detects low-probability consumption events that could signify faults, inefficiencies, or external intrusions. Experimental evaluation on the AEP dataset demonstrates that the proposed GMM framework achieves strong anomaly detection performance, with a precision of 0.89, recall of 0.84, and ROC-AUC of 0.91. Comparative results with threshold-based detection, k-means clustering, isolation forest, LSTM autoencoders, and fuzzy c-means highlight the superiority of GMM in capturing multimodal usage patterns. These findings suggest that the proposed adaptive GMM approach offers a reliable and interpretable tool for real-time anomaly detection in smart household grids.
