Scholarlink

<p>When harvesting daylilies, insufficient picking accuracy often occurs due to dense plant occlusion and difficult category differentiation. To solve this problem, we propose a localization method based on AD-MDNet (Adaptive Dual-branch MW-GCA Daylily Network). We design a dual-branch decoder fusing local details and global semantics, construct a progressive ladder feature miner module with multi-receptive fields and multi-resolution features, a pyramid-class enhancement network module for multilevel daylily classification, and propose a parallel artificial lemming optimization algorithm to optimize the hyperparameters of deep neural networks. These components constitute ADLCE-Net (Adaptive Dual-branch Loss and Category Enhancement Network). Based on segmentation results, the MW-GCA (Mini-Window Guided Corner Analysis) algorithm for intelligent picking localization is designed: it screens corner coordinates, extracts plant posture line segments via a U-shaped local window, integrates depth camera–based 3D modeling to establish the 3D coordinates of picking points, and achieves accurate estimation of their pixel coordinates and posture angles. Experimental results on the self-constructed TYUST-Daylily dataset show that, compared with comparative methods, ADLCE-Net reduces the number of parameters by an average of 62.3%, with average improvements of 4.89% and 4.23% in mIoU and mPA, respectively, verifying the effectiveness and reliability of the proposed method.</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

Research on method for dense daylily harvesting point localization based on AD-MDNet

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Rui Zhang,
  • Xiaonan Wei,
  • Yanjun Zhang,
  • Qian Wang,
  • Haihong Li

摘要

When harvesting daylilies, insufficient picking accuracy often occurs due to dense plant occlusion and difficult category differentiation. To solve this problem, we propose a localization method based on AD-MDNet (Adaptive Dual-branch MW-GCA Daylily Network). We design a dual-branch decoder fusing local details and global semantics, construct a progressive ladder feature miner module with multi-receptive fields and multi-resolution features, a pyramid-class enhancement network module for multilevel daylily classification, and propose a parallel artificial lemming optimization algorithm to optimize the hyperparameters of deep neural networks. These components constitute ADLCE-Net (Adaptive Dual-branch Loss and Category Enhancement Network). Based on segmentation results, the MW-GCA (Mini-Window Guided Corner Analysis) algorithm for intelligent picking localization is designed: it screens corner coordinates, extracts plant posture line segments via a U-shaped local window, integrates depth camera–based 3D modeling to establish the 3D coordinates of picking points, and achieves accurate estimation of their pixel coordinates and posture angles. Experimental results on the self-constructed TYUST-Daylily dataset show that, compared with comparative methods, ADLCE-Net reduces the number of parameters by an average of 62.3%, with average improvements of 4.89% and 4.23% in mIoU and mPA, respectively, verifying the effectiveness and reliability of the proposed method.