Combining Multi-Scale Fusion and Attentional Mechanisms for Assessing Writing Accuracy

Applied Sciences, Journal Year: 2025, Volume and Issue: 15(3), P. 1204 - 1204

Published: Jan. 24, 2025

Traditional methods of assessing handwritten characters are often too subjective, inefficient, and lagging in feedback, which makes it difficult for educators to achieve fully objective writing assessments writers receive timely suggestions improvement. In this paper, we propose a convolutional neural network (CNN) architecture that combines the attention mechanism with multi-scale feature fusion; specifically, features weighted by designing bottleneck layer Squeeze-and-Excitation (SE) highlight important information applying fusion method enable capture both global structure local details Chinese characters. Finally, high-quality dataset containing 26,800 images is constructed based on application scenario grade test, covering common exam; The experimental results show proposed achieves 98.6% accuracy exam 97.05% ICDAR-2013 public dataset, significantly improving recognition accuracy. improved model suitable scenarios such as exams, helps improve marking efficiency

Language: Английский

---

YOLO-SG: Seafloor Topography Unit Recognition and Segmentation Algorithm Based on Lightweight Upsampling Operator and Attention Mechanisms

Journal of Marine Science and Engineering, Journal Year: 2025, Volume and Issue: 13(3), P. 583 - 583

Published: March 16, 2025

The recognition and segmentation of seafloor topography play a crucial role in marine science research engineering applications. However, traditional methods for face several issues, such as poor capability analyzing complex terrains limited generalization ability. To address these challenges, this study introduces the SG-MKD dataset (Submarine Geomorphology Dataset—Seamounts, Sea Knolls, Submarine Depressions) proposes YOLO-SG (You Only Look Once—Submarine Geomorphology), an algorithm topographic unit that leverages lightweight upsampling operator attention mechanisms. provides instance annotations three types units—seamounts, sea knolls, submarine depressions—across total 419 images. is optimized version YOLOv8l-Segment model, incorporating convolutional block module backbone network to enhance feature extraction. Additionally, it integrates lightweight, general create new fusion network, thereby improving model’s ability fuse represent features. Experimental results demonstrate significantly outperforms original YOLOv8l-Segment, with 14.7% increase mean average precision. Furthermore, inference experiments conducted across various areas highlight strong capability.

Language: Английский

---

Depth Estimation Based on MMwave Radar and Camera Fusion with Attention Mechanisms and Multi-Scale Features for Autonomous Driving Vehicles

Electronics, Journal Year: 2025, Volume and Issue: 14(2), P. 300 - 300

Published: Jan. 13, 2025

Autonomous driving vehicles have strong path planning and obstacle avoidance capabilities, which provide great support to avoid traffic accidents. has become a research hotspot worldwide. Depth estimation is key technology in autonomous as it provides an important basis for accurately detecting objects avoiding collisions advance. However, the current difficulties depth include insufficient accuracy, difficulty acquiring information using monocular vision, challenge of fusing multiple sensors estimation. To enhance performance complex environments, this study proposes method point clouds images obtained from MMwave radar cameras are fused. Firstly, residual network established extract multi-scale features corresponding image simultaneously same location. Correlations between points by extracted features. A semi-dense achieved assigning value most relevant region. Secondly, bidirectional feature fusion structure with additional branches designed richness information. The loss during process reduced, robustness model enhanced. Finally, parallel channel position attention mechanisms used representation areas fused map, interference irrelevant suppressed, accuracy experimental results on public dataset nuScenes show that, compared baseline model, proposed reduces average absolute error (MAE) 4.7–6.3% root mean square (RMSE) 4.2–5.2%.

Language: Английский

---

Robust Momentum-Enhanced Non-Negative Tensor Factorization for Accurate Reconstruction of Incomplete Power Consumption Data

Electronics, Journal Year: 2025, Volume and Issue: 14(2), P. 351 - 351

Published: Jan. 17, 2025

Power consumption (PC) data are fundamental for optimizing energy use and managing industrial operations. However, with the widespread adoption of data-driven technologies in sector, maintaining integrity quality these has become a significant challenge. Missing or incomplete data, often caused by equipment failures communication disruptions, can severely affect accuracy reliability analyses, ultimately leading to poor decision-making increased operational costs. To address this, we propose Robust Momentum-Enhanced Non-Negative Tensor Factorization (RMNTF) model, which integrates three key innovations. First, model utilizes adversarial loss L2 regularization enhance its robustness improve performance when dealing data. Second, sigmoid function is employed ensure that results remain non-negative, aligning inherent characteristics PC improving analysis. Finally, momentum optimization applied accelerate convergence process, significantly reducing computational time. Experiments conducted on two publicly available datasets, densities 6.65% 4.80%, show RMNTF outperforms state-of-the-art methods, achieving an average reduction 16.20% imputation errors improvement 68.36% efficiency. These highlight model’s effectiveness handling sparse ensuring reconstructed support critical tasks like optimization, smart grid maintenance, predictive analytics.

Language: Английский

---

A Graph Neural Network-Based Context-Aware Framework for Sentiment Analysis Classification in Chinese Microblogs

Mathematics, Journal Year: 2025, Volume and Issue: 13(6), P. 997 - 997

Published: March 18, 2025

Sentiment analysis in Chinese microblogs is challenged by complex syntactic structures and fine-grained sentiment shifts. To address these challenges, a Contextually Enriched Graph Neural Network (CE-GNN) proposed, integrating self-supervised learning, context-aware embeddings, Networks (GNNs) to enhance classification. First, CE-GNN pre-trained on large corpus of unlabeled text through where Masked Language Modeling (MLM) Next Sentence Prediction (NSP) are leveraged obtain contextualized embeddings. These embeddings then refined embedding layer, which dynamically adjusted based the surrounding improve sensitivity. Next, dependencies captured (GNNs), words represented as nodes relationships denoted edges. Through this graph-based structure, sentence structures, particularly Chinese, can be interpreted more effectively. Finally, model fine-tuned labeled dataset, achieving state-of-the-art performance Experimental results demonstrate that achieves superior accuracy, with Macro F-measure 80.21% Micro 82.93%. Ablation studies further confirm each module contributes significantly overall performance.

Language: Английский

---