Scholarlink

<p><i>Brassica oleracea</i> var. <i>botrytis</i> is susceptible to bacterial spot, black rot, and downy mildew, causing yield loss and quality decline. Existing single-modal RGB detection is sensitive to illumination and imaging conditions, prone to feature ambiguity and overfitting, and lacks robustness. Depth imaging can characterize the three-dimensional morphology and spatial distribution of crops, providing geometric and structural information, which compensates for the limitations of visible light in boundary and morphology differentiation. To address this, this paper proposes a hierarchical multimodal fusion network (HMFNet), which adopts dual-path feature extraction to achieve differential decoupling and refinement of RGB and depth information, integrating multiscale complementary information in a two-stage fusion process. Within this framework, three key modules are proposed: the Cross-Modal Complementary Boundary Aggregation (CCBA) module, which combines multiscale feature selection and explicit edge enhancement to improve boundary separability; the Cross-Stage Partial with Adaptive Self-Attention (C2ASA) module, which enhances global representation through adaptive feature modulation and long-range dependency modeling; and the Wavelet-Guided Attention (WGA) module, which utilizes frequency-domain decomposition to achieve differential modeling of high- and low-frequency components, thereby suppressing complex background noise and highlighting discriminative regions. Experimental results demonstrate that HMFNet achieves an average precision (P), recall (R), F1-score, and mAP@0.5 of 93.2%, 87.9%, 90.3%, and 92%, respectively, on a dataset encompassing bacterial spot, black rot, downy mildew, and healthy. The mAP@0.5 for three disease categories reaches 90.6%, 85.0%, and 93.0%, respectively, significantly outperforming mainstream methods such as YOLO and RT-DETR, thus verifying the method’s effectiveness and practical value in complex field environments.</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

Integration of boundary enhancement and wavelet-guided attention in a hierarchical multimodal network for Brassica oleracea var. botrytis disease detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月17日

查看全文
  • Jianxin Xue,
  • Tianzuo Li,
  • Peizeng Wu,
  • Zhenyu Xie,
  • Zhiqi Liu,
  • Miaomiao Wei,
  • Xinming Yuan,
  • Xindong Wang,
  • Zimeng Zhang

摘要

Brassica oleracea var. botrytis is susceptible to bacterial spot, black rot, and downy mildew, causing yield loss and quality decline. Existing single-modal RGB detection is sensitive to illumination and imaging conditions, prone to feature ambiguity and overfitting, and lacks robustness. Depth imaging can characterize the three-dimensional morphology and spatial distribution of crops, providing geometric and structural information, which compensates for the limitations of visible light in boundary and morphology differentiation. To address this, this paper proposes a hierarchical multimodal fusion network (HMFNet), which adopts dual-path feature extraction to achieve differential decoupling and refinement of RGB and depth information, integrating multiscale complementary information in a two-stage fusion process. Within this framework, three key modules are proposed: the Cross-Modal Complementary Boundary Aggregation (CCBA) module, which combines multiscale feature selection and explicit edge enhancement to improve boundary separability; the Cross-Stage Partial with Adaptive Self-Attention (C2ASA) module, which enhances global representation through adaptive feature modulation and long-range dependency modeling; and the Wavelet-Guided Attention (WGA) module, which utilizes frequency-domain decomposition to achieve differential modeling of high- and low-frequency components, thereby suppressing complex background noise and highlighting discriminative regions. Experimental results demonstrate that HMFNet achieves an average precision (P), recall (R), F1-score, and mAP@0.5 of 93.2%, 87.9%, 90.3%, and 92%, respectively, on a dataset encompassing bacterial spot, black rot, downy mildew, and healthy. The mAP@0.5 for three disease categories reaches 90.6%, 85.0%, and 93.0%, respectively, significantly outperforming mainstream methods such as YOLO and RT-DETR, thus verifying the method’s effectiveness and practical value in complex field environments.