Scholarlink

<p>Addressing the need for a trade-off among accuracy, efficiency, and model size in remote sensing image segmentation, we propose MDAB-UNet: an efficient segmentation network incorporating Multiscale Dual Attention and a Soft Attention Bottleneck Convolution. Building upon the VGG-UNet architecture, MDAB-UNet introduces three core improvements: First, a Soft Attention Bottleneck Module (SABM) reduces parameters and computational cost while preserving representational capacity through soft gating and improved gradient flow. Second, a Multi-scale Dual Attention (MSDA) skip connection fuses multi-level features and employs dilated convolutions to expand receptive fields, enhancing detail recovery and contextual perception for objects of varying scales. Finally, the encoder–decoder structure is optimized by incorporating depthwise separable convolutions, unifying convolution counts, adding attention mechanisms, and residual connections, further balancing performance and computational burden. Experiments on two remote sensing datasets, LoveDA and WHDLD, demonstrate that MDAB-UNet achieves 48.36% and 62.21% mIoU, respectively, with 11.64 M parameters and 64.66 G FLOPs. Edge device tests further validate its deployment potential: with TensorRT acceleration, MDAB-UNet achieves real-time inference exceeding 26 FPS on the Jetson AGX Orin platform. Compared to recent lightweight segmentation models, the proposed method offers a favorable trade-off among parameter count, computational complexity, and accuracy, making it suitable for resource-constrained remote sensing applications.</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

MDAB-UNet: a lightweight and efficient improved UNet architecture for remote sensing image segmentation

  • 期刊论文

  • 发表日期：

    2026年5月18日

查看全文
  • Yunhui Zhang,
  • Xiaoxu Wei,
  • Yongsheng Wang

摘要

Addressing the need for a trade-off among accuracy, efficiency, and model size in remote sensing image segmentation, we propose MDAB-UNet: an efficient segmentation network incorporating Multiscale Dual Attention and a Soft Attention Bottleneck Convolution. Building upon the VGG-UNet architecture, MDAB-UNet introduces three core improvements: First, a Soft Attention Bottleneck Module (SABM) reduces parameters and computational cost while preserving representational capacity through soft gating and improved gradient flow. Second, a Multi-scale Dual Attention (MSDA) skip connection fuses multi-level features and employs dilated convolutions to expand receptive fields, enhancing detail recovery and contextual perception for objects of varying scales. Finally, the encoder–decoder structure is optimized by incorporating depthwise separable convolutions, unifying convolution counts, adding attention mechanisms, and residual connections, further balancing performance and computational burden. Experiments on two remote sensing datasets, LoveDA and WHDLD, demonstrate that MDAB-UNet achieves 48.36% and 62.21% mIoU, respectively, with 11.64 M parameters and 64.66 G FLOPs. Edge device tests further validate its deployment potential: with TensorRT acceleration, MDAB-UNet achieves real-time inference exceeding 26 FPS on the Jetson AGX Orin platform. Compared to recent lightweight segmentation models, the proposed method offers a favorable trade-off among parameter count, computational complexity, and accuracy, making it suitable for resource-constrained remote sensing applications.