Scholarlink

<p>Traditional deep learning-based algorithms for ship radiated noise recognition struggle to balance accuracy and network complexity. To address this, we propose MobileViNeXt, a lightweight fusion network that achieves both high recognition accuracy and efficient implementation. The network employs a SoftPool module to reduce redundant features in MobileViT, introduces a GhostNet-inspired inverted residual block to enhance feature richness, and replaces the multi-head self-attention mechanism with a decoupled fully connected module to lower complexity. To enhance the robustness of the model, a cross-shortcut connection integrates the MobileViXt backbone with a MobileNeXt branch to fuse global and fine-grained features. Experimental results show that MobileViNeXt achieves 98.1% and 98.5% accuracy on the ShipsEar and DeepShip datasets, respectively, with only 2.3&#xa0;million parameters and 0.98G FLOPs. Notably, it improves accuracy by 1.7% over the lightweight CFTANet and reduces parameters by 85% compared to the high-accuracy MSLEFC, demonstrating an excellent balance between performance and deploy ability for resource-constrained sonar systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MobileViNeXt: a lightweight fusion model for ship-radiated noise recognition

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Ming Chen,
  • Feng Wang,
  • Chaofeng Cao,
  • Hongwei Chen,
  • Yangze Dong

摘要

Traditional deep learning-based algorithms for ship radiated noise recognition struggle to balance accuracy and network complexity. To address this, we propose MobileViNeXt, a lightweight fusion network that achieves both high recognition accuracy and efficient implementation. The network employs a SoftPool module to reduce redundant features in MobileViT, introduces a GhostNet-inspired inverted residual block to enhance feature richness, and replaces the multi-head self-attention mechanism with a decoupled fully connected module to lower complexity. To enhance the robustness of the model, a cross-shortcut connection integrates the MobileViXt backbone with a MobileNeXt branch to fuse global and fine-grained features. Experimental results show that MobileViNeXt achieves 98.1% and 98.5% accuracy on the ShipsEar and DeepShip datasets, respectively, with only 2.3 million parameters and 0.98G FLOPs. Notably, it improves accuracy by 1.7% over the lightweight CFTANet and reduces parameters by 85% compared to the high-accuracy MSLEFC, demonstrating an excellent balance between performance and deploy ability for resource-constrained sonar systems.