Global Journal of Multidisciplinary and Applied Sciences

Monaural Speech Enhancement with Selective Local and Non-Local Attention

Lea Girard — 2026-05-10

Monaural speech enhancement remains a formidable challenge in audio signal processing, primarily due to the absence of spatial cues that typically facilitate the separation of target speech from background interference. Recent advancements in deep learning have significantly improved the quality and intelligibility of enhanced speech, yet balancing the extraction of fine-grained local acoustic features with the comprehension of global contextual dependencies remains an ongoing dilemma. This paper presents a novel framework that integrates a selective local and non-local attention mechanism to dynamically model both short-term phonetic characteristics and long-term acoustic environments. The local attention module focuses on preserving transient speech components and preserving high-frequency details, while the non-local attention mechanism captures long-range dependencies, aiding in the suppression of stationary and non-stationary noises over extended temporal receptive fields. Furthermore, a selective gating mechanism is introduced to adaptively fuse the outputs of these two attention branches, allocating computational focus based on the instantaneous characteristics of the input signal. Comprehensive evaluations on standard benchmark datasets demonstrate that the proposed architecture achieves state-of-the-art performance across multiple objective metrics, including perceptual evaluation of speech quality and short-time objective intelligibility. The results indicate that the dynamic fusion of local and global contexts significantly mitigates speech distortion and noise residual artifacts, particularly in low signal-to-noise ratio conditions.

Transfer Learning and Generative Modeling for Low-Resource Language Processing: Recent Advances

Jonathan A Smith, — 2026-04-15

The rapid evolution of natural language processing has predominantly benefited a small subset of the worlds languages, leaving the vast majority underrepresented in the digital era. This paper provides a comprehensive analysis of recent advancements in addressing this linguistic inequality through the dual lenses of transfer learning and generative modeling. We systematically explore how cross-lingual transfer mechanisms enable the projection of learned representations from resource-rich domains to low-resource targets, mitigating the fundamental challenge of data sparsity. Furthermore, we investigate the paradigm shift introduced by large generative models, which possess unprecedented capabilities for synthetic data augmentation, zero-shot inference, and few-shot adaptation. By synthesizing theoretical frameworks and empirical observations, we evaluate the efficacy of parameter-efficient fine-tuning techniques, typologically informed transfer strategies, and prompt-based learning methodologies. Our analysis highlights the intersection of linguistic typology and machine learning architectures, demonstrating that structural similarities between source and target languages significantly dictate the success of representation alignment. Finally, we address the critical limitations inherent in current approaches, including the amplification of algorithmic bias, the phenomena of negative transfer, and the challenges associated with the subword tokenization of morphologically rich languages. The insights presented herein aim to guide future research toward more equitable and robust multilingual systems.

An Integrated Framework for Branch Detection and Depth Estimation in UAV Stereo Vision for Forestry Pruning

Sofia Keller, — 2026-04-01

The automation of forestry management practices, particularly selective branch pruning, represents a significant challenge in modern silviculture. Manual pruning is labor-intensive, time-consuming, and presents considerable safety risks to human operators. While Unmanned Aerial Vehicles have been extensively deployed for passive remote sensing and canopy analysis, their application in active physical interaction tasks such as pruning remains limited by the complexities of aerial manipulation in unstructured environments. A critical prerequisite for autonomous aerial pruning is the precise visual identification and spatial localization of target branches. This paper proposes a comprehensive and integrated framework that seamlessly combines deep learning based semantic segmentation for robust branch detection with binocular stereo vision for high accuracy depth estimation. The proposed system is designed to operate onboard a resource constrained Unmanned Aerial Vehicle, processing complex canopy imagery to output isolated three-dimensional branch coordinates suitable for guiding a robotic pruning effector. By integrating a lightweight convolutional neural network with a highly optimized semi-global stereo matching algorithm, the framework addresses the inherent challenges of dynamic lighting, heavy visual occlusion, and background clutter characteristic of forest environments. Extensive field experiments and mock-up trials demonstrate the efficacy of the proposed pipeline. The semantic segmentation module achieves high pixel wise accuracy in isolating branch structures from surrounding foliage, while the stereo vision component provides reliable depth maps with a minimal margin of error. The synthesized spatial data allows for the accurate extraction of branch cutting points. This research contributes a crucial foundational technology toward the realization of fully autonomous aerial forestry tools, bridging the gap between passive observation and active robotic intervention in complex natural landscapes.

Transformer-Based Spatial-Temporal Models for Comprehensive Scene Understanding Object Tracking and Autonomous Decision Support

Amelia Paredes, — 2026-04-05

The integration of scene understanding, object tracking, and decision support into a singular computational framework remains a formidable challenge in autonomous systems. Traditional approaches have relied on disjointed pipelines where convolutional neural networks process spatial features, recursive algorithms manage temporal tracking, and isolated heuristic models handle downstream decision making. Such fragmentation inherently introduces cascading errors, latency, and suboptimal context sharing. In this paper, we propose a unified Transformer-based architecture designed to concurrently process spatial-temporal representations for holistic scene understanding, continuous target tracking, and proactive decision support. By leveraging self-attention mechanisms across both spatial dimensions and temporal frames, the proposed model efficiently constructs global contextual dependencies without the restricted receptive fields characteristic of conventional convolutions. Our methodology incorporates a multi-head prediction module that projects shared latent embeddings into semantic segmentation masks, object bounding boxes, and action policy probabilities. We conduct extensive empirical evaluations on standard large-scale driving datasets, demonstrating that our integrated spatiotemporal Transformer significantly reduces inference latency while achieving superior quantitative metrics across all three domains compared to state-of-the-art disjointed architectures. The findings underscore the efficacy of global representation learning in complex dynamic environments and provide a robust foundation for the next generation of autonomous robotic and vehicular control systems.

Methods for Enhancing Factuality of Large Language Models via Retrieval-Augmented Mechanisms

Julian Sterling, — 2026-04-10

The rapid proliferation of large language models has fundamentally transformed the landscape of natural language processing, enabling unprecedented capabilities in text generation, summarization, and interactive dialogue. However, a persistent and critical limitation of these generative architectures is their propensity to produce factually incorrect or unverified information, a phenomenon widely characterized as hallucination. This paper presents a comprehensive investigation into methods for mitigating hallucinatory behaviors and enhancing the factuality of large language models through the implementation of advanced retrieval-augmented mechanisms. By dynamically decoupling the parametric memory of the neural network from a non-parametric, externally updatable knowledge base, retrieval-augmented generation paradigms offer a robust solution to the limitations of static pre-training. We provide a deep architectural analysis of the integration between dense passage retrieval systems and autoregressive generation processes. Furthermore, we propose a novel contextual attention mechanism designed to optimize the semantic fusion of retrieved documents with user prompts. Through extensive empirical evaluations on standard knowledge-intensive datasets, we demonstrate that our refined retrieval-augmented framework significantly outperforms conventional parametric baselines and standard heuristic retrieval approaches. The results indicate substantial improvements in exact match metrics and a dramatic reduction in hallucination rates. This research elucidates the theoretical underpinnings of factuality in generative models and establishes a scalable, algorithmically efficient framework for deploying highly reliable artificial intelligence systems in mission-critical applications.

Photorealistic Video Colorization Using Gated Color Guidance and Cross-Frame Consistency

Nozomi Okada — 2026-05-10

Video colorization remains a profoundly challenging problem in the domain of computer vision, demanding not only accurate spatial colorization but also robust temporal consistency across sequential frames. Previous approaches frequently suffer from severe visual artifacts, notably color bleeding, temporal flickering, and semantic mismatch, which collectively degrade the photorealism of the resulting outputs. To mitigate these pervasive issues, this paper introduces a novel framework for photorealistic video colorization utilizing gated color guidance alongside an advanced cross-frame consistency mechanism. The gated color guidance module effectively selectively incorporates prior color information from exemplar frames, dynamically weighing the relevance of reference colors based on deep semantic features. Concurrently, the cross-frame consistency module employs recurrent feature propagation to ensure that temporal variations remain imperceptible to the human visual system, thereby effectively eliminating flickering artifacts. Through rigorous experimental evaluation on standard benchmark datasets, the proposed architecture demonstrates unprecedented performance improvements across various quantitative metrics and qualitative visual assessments. The ablation studies validate the critical contributions of both the gating mechanism and the temporal consistency regularization. This research establishes a robust foundation for future applications in film restoration, historical archive digitization, and automated video enhancement.

Text-to-SQL Agents Under Ambiguous User Intent: A Taxonomy, Benchmark, and Repair Strategy

Giulia Kruger — 2026-05-10

The translation of natural language queries into executable database queries, commonly known as Text-to-SQL, has seen remarkable progress with the advent of large language models. However, standard benchmarking frameworks implicitly assume that user queries are fully specified, structurally sound, and semantically unambiguous. In real-world enterprise deployments, user intents are frequently characterized by missing constraints, vague terminology, and structural ambiguity, leading autonomous agents to generate plausible but incorrect SQL queries. This paper presents a comprehensive investigation into the behavior of Text-to-SQL agents operating under conditions of ambiguous user intent. We introduce a novel, fine-grained taxonomy that categorizes linguistic and structural ambiguities specific to relational database querying. To empirically evaluate agent performance, we propose a new evaluation benchmark comprising thousands of naturally ambiguous queries paired with multivalent target interpretations. Furthermore, we develop a conversational repair strategy that equips Text-to-SQL agents with the ability to detect ambiguity, formulate targeted clarification questions, and iteratively refine the generated queries based on user feedback. Through extensive experimental analysis, we demonstrate that current state-of-the-art models suffer severe performance degradation when exposed to ambiguous inputs. The implementation of our proposed interactive repair framework recovers a significant portion of this lost accuracy, reducing critical semantic errors while maintaining a low cognitive burden on the user.

Multi-Source Data Fusion for Perception in Agricultural and Forestry Scenarios: A Comprehensive Analysis

Noah Rossi — 2026-04-15

The automation of agricultural and forestry operations relies fundamentally on the capacity of autonomous systems to perceive and interpret highly unstructured, dynamic, and complex environments. Traditional perception systems relying on single-modality sensors, such as standalone optical cameras or isolated light detection and ranging systems, frequently encounter severe performance degradation when subjected to the harsh realities of these domains. These challenges include variable illumination, severe occlusion by dense foliage, atmospheric disturbances like dust and fog, and irregular terrain topologies. This paper provides a comprehensive analysis of multi-source data fusion methodologies tailored specifically for agricultural and forestry scenarios. By synergistically integrating data from vision sensors, light detection and ranging, and millimeter-wave radar, autonomous platforms can achieve a level of robust situational awareness previously unattainable. The research explores the underlying principles of spatial and temporal calibration across heterogeneous sensor suites and details advanced preprocessing techniques necessary for aligning disparate data modalities. Furthermore, the study evaluates hierarchical fusion architectures, encompassing data-level, feature-level, and decision-level integration strategies. The findings indicate that feature-level fusion, particularly when facilitated by deep learning frameworks such as cross-modality attention mechanisms, yields significant improvements in obstacle detection, terrain mapping, and crop phenotyping accuracy under degraded environmental conditions. Ultimately, this comprehensive review and analysis aim to establish a foundational framework for future developments in resilient autonomous perception systems across complex biological terrains.

Systemic Biomarkers and Chronic Disease Comorbidity in Age-Related Macular Degeneration

Lorenzo Bergmann — 2026-05-15

Age-related macular degeneration remains a leading cause of irreversible visual impairment among older adults globally. Historically conceptualized as a localized ocular pathology, emerging epidemiological and molecular evidence necessitates a paradigm shift towards understanding the condition as a localized manifestation of broader systemic dysfunction. This paper provides a comprehensive analysis of the systemic biomarkers bridging age-related macular degeneration with chronic systemic comorbidities, particularly cardiovascular disease, metabolic syndrome, and neurodegenerative disorders. By synthesizing current pathophysiological frameworks, we elucidate the shared biological mechanisms underlying these conditions, including chronic low-grade inflammation, lipid dysregulation, and complement cascade hyperactivation. We examine established and novel circulating biomarkers, detailing the methodological complexities inherent in their quantification and statistical modeling within large-scale epidemiological cohorts. The findings highlight the significant prognostic value of systemic inflammatory markers, such as C-reactive protein and various interleukins, in predicting both ocular disease progression and the onset of systemic morbidities. Furthermore, the convergence of pathogenic pathways suggests that patients presenting with specific retinal phenotypes should be considered for comprehensive systemic evaluations. Ultimately, this integration of ophthalmic and systemic clinical data advocates for a multidisciplinary approach to patient management, fostering the development of targeted, systemic therapeutic interventions capable of mitigating both visual decline and comorbid disease burden.

A Unified Framework for Deep Reconstruction Enhancement and Anomaly Detection

Amelia ODonnell, — 2026-04-15

Anomaly detection in high-dimensional data streams remains a fundamental challenge in computer science, particularly when deploying robust machine learning systems in unpredictable real-world environments. Traditional unsupervised methods often struggle with a pervasive trade-off between accurately reconstructing normal data patterns and inadvertently over-reconstructing anomalous instances, which fundamentally degrades the distinctiveness of the anomaly score. In this paper, we propose a comprehensive unified framework for deep reconstruction enhancement and anomaly detection that mitigates these pathological memorization effects while preserving high fidelity for in-distribution representations. Our architecture introduces a novel dual-pathway feature enhancement module integrated with a multi-scale autoencoding backbone, which structurally constrains the latent space manifold to isolate and amplify reconstruction errors specifically for anomalous perturbations. By explicitly formulating a joint optimization objective that simultaneously maximizes representation quality for normal instances and enforces tight bounding around the nominal manifold, our approach achieves exceptional discriminative power. We conduct extensive empirical evaluations across multiple complex domains, demonstrating superior performance in standard metrics such as the area under the receiver operating characteristic curve. The proposed system effectively bridges the gap between generative fidelity and diagnostic sensitivity, establishing a new operational standard for automated defect detection, network intrusion monitoring, and medical image screening.

Efficient Edge Video Analytics with Region-Aware Enhancement and Temporal Consistency

Jisoo Kang — 2026-05-25

The exponential proliferation of connected vision sensors has fundamentally transformed the landscape of automated surveillance, intelligent transportation systems, and industrial monitoring. Conventional paradigms that rely on transmitting continuous, high-definition video streams to centralized cloud architectures are increasingly untenable due to severe bandwidth constraints, inherent transmission latency, and profound privacy concerns. Edge computing has emerged as a compelling alternative by migrating computational resources closer to the data source. However, edge devices frequently possess constrained computational capabilities and limited thermal budgets, rendering the execution of complex deep neural networks highly challenging. This research presents a comprehensive framework for efficient edge video analytics characterized by two novel components. First, we introduce a region-aware enhancement mechanism that selectively allocates computational resources to spatial areas of high analytical value, thereby discarding irrelevant background information and significantly reducing spatial redundancy. Second, we integrate a temporal consistency module designed to leverage the inherent continuity across sequential frames. By propagating high-level semantic features from previous frames to current frames using lightweight motion estimation, the system minimizes redundant computations while ensuring smooth and stable analytical outputs. Through extensive evaluation on standard video analytic datasets, our proposed methodology demonstrates substantial improvements in processing speed and bandwidth utilization without compromising analytical accuracy.

Steel-Glass Structural Analysis for Architectural Design Under Global Stability Constraints

Sebastien Arnaud — 2026-05-01

The integration of steel and glass in modern architectural design has fundamentally transformed the aesthetic and functional capabilities of contemporary building envelopes and load-bearing structures. While transparency and minimal visual obstruction remain primary design drivers, the structural interplay between high-strength steel frameworks and brittle glass panels introduces significant complexities, particularly regarding global stability constraints. This paper presents a comprehensive analytical framework for evaluating steel-glass composite structures, focusing on the mitigation of buckling and the enhancement of overall system stability under variable environmental and operational loads. By treating glass not merely as cladding but as an active participant in structural load transfer, the analysis explores the synergistic stiffening effects that panels can provide to slender steel sub-structures. The research investigates material interaction, connection stiffness, and geometric nonlinearities to establish robust design protocols that satisfy stringent architectural stability requirements. Through advanced computational simulation strategies evaluated qualitatively, the study elucidates the influence of varying support conditions and load distributions on the global buckling behavior of composite assemblies. The findings offer critical insights into optimizing connection typologies and member proportions, ensuring that structural safety does not compromise architectural intent. The proposed analytical paradigms aim to bridge the gap between architectural vision and engineering reliability, providing designers with theoretical foundations for the safe implementation of expansive structural glass systems.

Reinforcement Learning-Controlled Ensemble Sampling for Imbalanced Complex Data

Alastair Cathrin — 2026-05-05

The proliferation of complex and highly imbalanced datasets across various domains poses a significant challenge to traditional machine learning algorithms, which inherently bias their decision boundaries toward majority classes. Existing data-level preprocessing techniques, such as static oversampling and undersampling, often fail to capture the underlying manifold structure of heterogeneous data, leading to problems such as severe overfitting, loss of crucial information, and degraded generalization capabilities. This paper introduces a novel framework utilizing a reinforcement learning agent to dynamically control ensemble sampling strategies tailored for imbalanced complex data. By modeling the sampling process as a Markov Decision Process, the proposed framework allows the agent to continuously interact with the data environment, observing local data distributions and iteratively adjusting the sampling ratios and synthetic generation parameters for multiple base classifiers. The reward mechanism is meticulously engineered to optimize global evaluation metrics, specifically emphasizing the minority class predictive performance without sacrificing majority class accuracy. Extensive theoretical analysis demonstrates how the reinforcement learning paradigm overcomes the rigid heuristics of conventional ensemble methods. Experimental validation on diverse, high-dimensional datasets confirms that the proposed approach substantially mitigates the adverse effects of extreme class overlap and noise. The findings suggest a paradigm shift in how sampling configurations are optimized, offering a robust, automated solution for complex predictive modeling tasks across interdisciplinary applications.