863 papers
This paper proposes a novel lightweight convolutional Transformer model utilizing multi-teacher hierarchical progressive knowledge distillation and directional compression to achieve effective surveillance video compression and quality enhancement while maintaining high accuracy and low memory consumption.
This paper introduces PV-YOLO, a lightweight, collaborative improvement of YOLOv8n featuring four custom modules that significantly enhance micro-defect detection accuracy and recall on photovoltaic cells while reducing computational overhead, enabling real-time edge inspection at 140 FPS on UAVs.
This paper presents the first machine-checked formalization in Isabelle/HOL of the BMSSP recurrence underlying the 2025 deterministic O(m log²/³ n) SSSP algorithm, providing a non-vacuous, size-parametric proof of its O(|V|·(ln|V|)²/³) runtime on an unbounded graph family without relying on axioms or unproven assumptions.
This study demonstrates that EfficientNet-B0 outperforms other deep learning architectures, including Vision Transformers, in industrial character classification with restricted alphabets and scarce data, highlighting the continued advantage of convolutional inductive biases in such constrained environments.
This study demonstrates that a modular consortium blockchain framework integrating differential privacy, homomorphic encryption, and smart contracts achieves clinical equivalence to standard federated learning for hospital mortality and radiography tasks, proving that robust privacy and security measures do not compromise predictive utility.
This paper introduces "collapse distance," a configurable safety-margin metric for quantifying network failure proximity, and proposes TCR-GIN, a physics-informed graph learning framework that efficiently and accurately estimates this metric to enable real-time safety management and early-warning signals across diverse complex networks.
This paper proposes AVTF-Net, a multimodal deep learning framework that integrates a modified AlexNet and fine-tuned BERT with an early fusion and attention mechanism to effectively detect fake news by capturing cross-modal inconsistencies, achieving state-of-the-art performance with 95.80% accuracy on the Fakeddit dataset.
This paper proposes a lightweight, resource-efficient post-OCR correction pipeline specifically designed to improve the accuracy of Vietnamese scanned documents by addressing diacritic errors through a combination of image preprocessing, dictionary-based correction, and rule-based restoration, offering a practical baseline for low-resource digitization scenarios.
This study employs a candidate-centric framework to demonstrate that trust is the strongest predictor of job applicants' intention to use AI-driven talent acquisition systems, significantly mediating the impact of ethical concerns like fairness, transparency, and privacy within the Indian context.
This paper presents a resource-efficient fine-tuning methodology using Low-Rank Adaptation (LoRA) and quantization to successfully deploy large Vision-Language-Action models on affordable, consumer-grade robotic platforms, demonstrating effective real-world manipulation performance with limited data.