SGK-Net: A Novel Navigation Scene Graph Generation Network DOI Creative Commons
Wenbin Yang, Hao Qiu, Xiangfeng Luo

et al.

Sensors, Journal Year: 2024, Volume and Issue: 24(13), P. 4329 - 4329

Published: July 3, 2024

Scene graphs can enhance the understanding capability of intelligent ships in navigation scenes. However, complex entity relationships and presence significant noise contextual information within scenes pose challenges for scene graph generation (NSGG). To address these issues, this paper proposes a novel NSGG network named SGK-Net. This comprises three innovative modules. The Semantic-Guided Multimodal Fusion (SGMF) module utilizes prior on relationship semantics to fuse multimodal construct features, thereby elucidating between entities reducing semantic ambiguity caused by relationships. Graph Structure Learning-based Evolution (GSLSE) module, based structure learning, reduces redundancy features optimizes computational complexity subsequent message passing. Key Entity Message Passing (KEMP) takes full advantage refine interference from non-key nodes. Furthermore, constructs first Ship Navigation Simulation dataset, SNSG-Sim, which provides foundational dataset research ship SGG. Experimental results SNSG-sim demonstrate that our method achieves an improvement 8.31% (R@50) PredCls task 7.94% SGCls compared baseline method, validating effectiveness generation.

Language: Английский

Digital Twin Model and Platform Based on a Dual System for Control Rod Drive Mechanism Safety DOI

Changfu Wan,

Wenqiang Li, Bo Yang

et al.

Reliability Engineering & System Safety, Journal Year: 2025, Volume and Issue: unknown, P. 111075 - 111075

Published: April 1, 2025

Language: Английский

Citations

0
Towards Super Compressed Neural Networks for Object Identification: Quantized Low-Rank Tensor Decomposition with Self-Attention DOI Open Access
Baichen Liu, Dongwei Wang, Qi Lv

et al.

Electronics, Journal Year: 2024, Volume and Issue: 13(7), P. 1330 - 1330

Published: April 2, 2024

Deep convolutional neural networks have a large number of parameters and require significant floating-point operations during computation, which limits their deployment in situations where the storage space is limited computational resources are insufficient, such as mobile phones small robots. Many network compression methods been proposed to address aforementioned issues, including pruning, low-rank decomposition, quantization, etc. However, these typically fail achieve ratio terms parameter count. Even when high rates achieved, network’s performance often significantly deteriorated, making it difficult perform tasks effectively. In this study, we propose more compact representation for networks, named Quantized Low-Rank Tensor Decomposition (QLTD), super compress deep networks. Firstly, employed Tucker decomposition pre-trained weights. Subsequently, further exploit redundancies within core tensor factor matrices obtained through vector quantization partition cluster Simultaneously, introduced self-attention module each matrix enhance training responsiveness critical regions. The object identification results CIFAR10 experiment showed that QLTD achieved 35.43×, with less than 1% loss accuracy 90.61×, 2% accuracy. was able count realize good balance between compressing maintaining

Language: Английский

Citations

1
Radar Perception of Multi-Object Collision Risk Neural Domains during Autonomous Driving DOI Open Access
Józef Lisowski

Electronics, Journal Year: 2024, Volume and Issue: 13(6), P. 1065 - 1065

Published: March 13, 2024

The analysis of the state literature in field methods perception and control movement autonomous vehicles shows possibilities improving them by using an artificial neural network to generate domains prohibited maneuvers passing objects, contributing increasing safety driving various real conditions surrounding environment. This article concerns radar perception, which involves receiving information about many then identifying assigning a collision risk preparing maneuvering response. In identification process, each object is assigned domain generated previously trained network. size proportional collisions distance changes during driving. Then, optimal trajectory determined from among possible safe paths, ensuring minimum time. presented solution task was illustrated with computer simulation situation objects. main achievements this are synthesis algorithm mapping objects characterizing their assessment degree on example multi-object simulation.

Language: Английский

Citations

0
SGK-Net: A Novel Navigation Scene Graph Generation Network DOI Creative Commons
Wenbin Yang, Hao Qiu, Xiangfeng Luo

et al.

Sensors, Journal Year: 2024, Volume and Issue: 24(13), P. 4329 - 4329

Published: July 3, 2024

Scene graphs can enhance the understanding capability of intelligent ships in navigation scenes. However, complex entity relationships and presence significant noise contextual information within scenes pose challenges for scene graph generation (NSGG). To address these issues, this paper proposes a novel NSGG network named SGK-Net. This comprises three innovative modules. The Semantic-Guided Multimodal Fusion (SGMF) module utilizes prior on relationship semantics to fuse multimodal construct features, thereby elucidating between entities reducing semantic ambiguity caused by relationships. Graph Structure Learning-based Evolution (GSLSE) module, based structure learning, reduces redundancy features optimizes computational complexity subsequent message passing. Key Entity Message Passing (KEMP) takes full advantage refine interference from non-key nodes. Furthermore, constructs first Ship Navigation Simulation dataset, SNSG-Sim, which provides foundational dataset research ship SGG. Experimental results SNSG-sim demonstrate that our method achieves an improvement 8.31% (R@50) PredCls task 7.94% SGCls compared baseline method, validating effectiveness generation.

Language: Английский

Citations

0