Scholarlink

<p>Accurate instance segmentation of tree species is crucial for forest ecosystem monitoring and biodiversity assessment. However, UAV-based forest imagery presents unique challenges due to the substantial variations in tree scales, intricate branch-leaf structures, and densely overlapping canopies. Recent query-based instance segmentation methods have achieved remarkable progress, but their direct application to forest scenes remains suboptimal. These models often fail to fully exploit cross-scale feature interactions and rely on randomly initialized query embeddings that lack image-aware priors. To address these issues, we proposed TreeSegNet, a query-based instance segmentation method built upon the Mask2Former framework, incorporating the following key modules: (1) Multi-scale Feature Segmentation Fusion module, which integrates multi-scale features via global feature fusion at the feature level and contextual fusion at the window level. (2) Multi-scale Frequency Mean Query Generator, which combines low-frequency information from multi-scale feature maps with content query vectors. (3) Gated-enhanced Queries Transformer Decoder, which selectively integrates query information from different levels by dynamically adjusting gating values. Experimental results show that TreeSegNet achieves significant improvements on a self-collected dataset of tree species instances, with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(AP_{50}\)</EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(AP_{75}\)</EquationSource> </InlineEquation> of 40.30% and 34.80%, respectively. It also performs competitively on the COCO and CVPPP leaf segmentation datasets, demonstrating strong generalization capability compared to state-of-the-art methods. Code and models of TreeSegNet are publicly available at <a href="https://github.com/FAFU-IMLab/TreeSegNet.">https://github.com/FAFU-IMLab/TreeSegNet.</a></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

TreeSegNet: multi-scale query-based instance segmentation with frequency-aware and gated feature enhancement

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Chih-Wei Lin,
  • Shangtai Zhou,
  • Lirong Zhu

摘要

Accurate instance segmentation of tree species is crucial for forest ecosystem monitoring and biodiversity assessment. However, UAV-based forest imagery presents unique challenges due to the substantial variations in tree scales, intricate branch-leaf structures, and densely overlapping canopies. Recent query-based instance segmentation methods have achieved remarkable progress, but their direct application to forest scenes remains suboptimal. These models often fail to fully exploit cross-scale feature interactions and rely on randomly initialized query embeddings that lack image-aware priors. To address these issues, we proposed TreeSegNet, a query-based instance segmentation method built upon the Mask2Former framework, incorporating the following key modules: (1) Multi-scale Feature Segmentation Fusion module, which integrates multi-scale features via global feature fusion at the feature level and contextual fusion at the window level. (2) Multi-scale Frequency Mean Query Generator, which combines low-frequency information from multi-scale feature maps with content query vectors. (3) Gated-enhanced Queries Transformer Decoder, which selectively integrates query information from different levels by dynamically adjusting gating values. Experimental results show that TreeSegNet achieves significant improvements on a self-collected dataset of tree species instances, with \(AP_{50}\) and \(AP_{75}\) of 40.30% and 34.80%, respectively. It also performs competitively on the COCO and CVPPP leaf segmentation datasets, demonstrating strong generalization capability compared to state-of-the-art methods. Code and models of TreeSegNet are publicly available at https://github.com/FAFU-IMLab/TreeSegNet.