Scholarlink

<p>To address the challenges in Unmanned Aerial Vehicle (UAV) infrared fire reconnaissance, such as background thermal interference, weak features of small targets, and the inadequacy of bounding boxes for fine-grained contour analysis, a lightweight instance segmentation model named Multi-scale, Sparse Transformer, and Lightweight YOLO (MSTL-YOLO) is proposed based on an improved YOLOv11n-seg. Specifically, the fundamental C3k2 building unit is reconstructed by integrating the Multi-Scale Edge Information Enhancement (MSEIE) design, utilizing a parallel structure to strengthen the perception of flame context and edge details. To bridge the semantic gap between scales, a Pyramid Sparse Transformer (PST) module is introduced to perform coarse-to-fine cross-scale semantic-detail interaction. Additionally, a Lightweight Shared Detail-Enhanced Convolution Decoding (LSDECD) head is employed, which realizes shared lightweight decoding with contour-aware enhancement to reduce parameters while ensuring high-quality mask generation. Experimental results show that the improved model achieves 93.3% and 90.1% on Box and Mask mAP@0.5, representing increases of 3.2% and 2.7% over the baseline, respectively. In addition, the model reduces parameters by 23.7% and GFLOPs by 12.7%, achieving an inference speed of 87 FPS on a GTX 1660 GPU. These results indicate that the proposed algorithm provides a favorable accuracy-efficiency trade-off and shows promise for edge-oriented UAV fire reconnaissance.</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

MSTL-YOLO: sparse transformer-enhanced lightweight instance segmentation for UAV infrared fire reconnaissance

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Sen Zhang,
  • Xinbo Chen,
  • Xiaojie Sun,
  • Lifan Sun,
  • Yue Wu

摘要

To address the challenges in Unmanned Aerial Vehicle (UAV) infrared fire reconnaissance, such as background thermal interference, weak features of small targets, and the inadequacy of bounding boxes for fine-grained contour analysis, a lightweight instance segmentation model named Multi-scale, Sparse Transformer, and Lightweight YOLO (MSTL-YOLO) is proposed based on an improved YOLOv11n-seg. Specifically, the fundamental C3k2 building unit is reconstructed by integrating the Multi-Scale Edge Information Enhancement (MSEIE) design, utilizing a parallel structure to strengthen the perception of flame context and edge details. To bridge the semantic gap between scales, a Pyramid Sparse Transformer (PST) module is introduced to perform coarse-to-fine cross-scale semantic-detail interaction. Additionally, a Lightweight Shared Detail-Enhanced Convolution Decoding (LSDECD) head is employed, which realizes shared lightweight decoding with contour-aware enhancement to reduce parameters while ensuring high-quality mask generation. Experimental results show that the improved model achieves 93.3% and 90.1% on Box and Mask mAP@0.5, representing increases of 3.2% and 2.7% over the baseline, respectively. In addition, the model reduces parameters by 23.7% and GFLOPs by 12.7%, achieving an inference speed of 87 FPS on a GTX 1660 GPU. These results indicate that the proposed algorithm provides a favorable accuracy-efficiency trade-off and shows promise for edge-oriented UAV fire reconnaissance.