Scholarlink

<p>The region of interest (ROI) is easily dominated by the background due to the imbalanced area distribution of feature weight information in the segmented images, which disturbs correct feature extraction. To address the segmentation challenges caused by the inherent morphological characteristics and imbalance feature weight information, we propose a novel network structure named TCDNet, which adopts <b>T</b>ransformer and <b>C</b>NN as its <b>D</b>ual-branch feature extractor to progressively sample the multi-scale channel semantic information while capturing both global and local information. We introduce a Shallow Fusion (SF) module to reconstruct channel dimensions and adaptively redistribute sampling feature weights. Additionally, to enhance feature fusion over long distances and facilitate interaction between layers, we design an improved upsampling structure called dense connection based on element-wise addition (DCEA). Furthermore, we implement a Feature Aggregation (FA) module to dynamically consolidate feature information and mitigate under-segmentation issues related to blurred boundary. This module employs an adaptive dual-branch pooling layer to integrate global feature information and incorporates an attention mechanism for dynamic weight redistribution. We evaluate segmentation experiments on Cityscapes dataset, LIDC-IDRI public dataset and in-house CT datasets. Compared to other methods, the proposed method significantly enhances the segmentation performance of the images and achieves promising results.</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

CNN and transformer feature adaptive fusion for image segmentation with imbalanced weight information

  • 期刊论文

  • 发表日期：

    2026年3月28日

查看全文
  • Hongxiang Peng,
  • Xinjun Bian,
  • Siyang Feng,
  • Jiale Chen,
  • Mingwei Chen,
  • Xipeng Pan,
  • Lingqiao Li,
  • Yang Zheng,
  • Ruojie Zhang

摘要

The region of interest (ROI) is easily dominated by the background due to the imbalanced area distribution of feature weight information in the segmented images, which disturbs correct feature extraction. To address the segmentation challenges caused by the inherent morphological characteristics and imbalance feature weight information, we propose a novel network structure named TCDNet, which adopts Transformer and CNN as its Dual-branch feature extractor to progressively sample the multi-scale channel semantic information while capturing both global and local information. We introduce a Shallow Fusion (SF) module to reconstruct channel dimensions and adaptively redistribute sampling feature weights. Additionally, to enhance feature fusion over long distances and facilitate interaction between layers, we design an improved upsampling structure called dense connection based on element-wise addition (DCEA). Furthermore, we implement a Feature Aggregation (FA) module to dynamically consolidate feature information and mitigate under-segmentation issues related to blurred boundary. This module employs an adaptive dual-branch pooling layer to integrate global feature information and incorporates an attention mechanism for dynamic weight redistribution. We evaluate segmentation experiments on Cityscapes dataset, LIDC-IDRI public dataset and in-house CT datasets. Compared to other methods, the proposed method significantly enhances the segmentation performance of the images and achieves promising results.