Scholarlink

<p>In rice breeding, improving both crop yield and quality is of paramount importance. This study investigates key factors directly influencing yield, particularly the number of rice panicles and leaf width. We hypothesize that increases in panicle count and leaf width correlate positively with photosynthetic efficiency, thereby significantly affecting overall crop yield. To test this hypothesis, we aim to evaluate and select rice varieties exhibiting desirable phenotypes through targeted detection techniques. We utilize an enhanced DINO (self-distillation with no labels) model for detecting and segmenting rice panicles and leaves. The upstream component of our model functions as an unsupervised general feature extractor, learning rich visual features from a large dataset of unlabeled rice images. The downstream task consists of two branches: one for detecting the number of rice panicles and another for segmenting leaf areas. By combining these two branches, we are able to accurately assess the photosynthetic potential and reproductive capacity of rice plants. Experimental results demonstrate that our model outperforms traditional methods in both panicle detection and leaf area segmentation, achieving higher accuracy and robustness. We conduct experiments on a newly curated dataset, RiceVar, which comprises over 50,000 images covering three rice cultivars captured under varied angles and backgrounds. Our proposed method achieves a mean average precision of 81.401 in panicle detection, a 14.7 point improvement over ResNet50, and a Dice coefficient of 84.322 and intersection over union of 82.186 in leaf segmentation, outperforming the EAPT model by 14.08 and 2.45 points, respectively. Moreover, our model remains stable under varying environmental conditions, highlighting its practical value for rice breeding applications. By precisely evaluating panicle count and leaf width, our model supports the selection of high-yield, high-efficiency rice varieties, contributing to the advancement of sustainable agricultural practices. The relevant code and data are available at <a href="https://github.com/xiaobeial/Semi-supervised-detection-and-segmentation-algorithm-for-efficient-rice-breeding.">https://github.com/xiaobeial/Semi-supervised-detection-and-segmentation-algorithm-for-efficient-rice-breeding.</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

Enhancing rice breeding efficiency through semi-supervised detection and segmentation of panicles and leaves

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月7日

查看全文
  • Yihong Hu,
  • Ling Xiong,
  • Bowen Chen,
  • Peiyi Yu,
  • Changrong Ye,
  • Gaofeng Jia,
  • Bingchuan Tian

摘要

In rice breeding, improving both crop yield and quality is of paramount importance. This study investigates key factors directly influencing yield, particularly the number of rice panicles and leaf width. We hypothesize that increases in panicle count and leaf width correlate positively with photosynthetic efficiency, thereby significantly affecting overall crop yield. To test this hypothesis, we aim to evaluate and select rice varieties exhibiting desirable phenotypes through targeted detection techniques. We utilize an enhanced DINO (self-distillation with no labels) model for detecting and segmenting rice panicles and leaves. The upstream component of our model functions as an unsupervised general feature extractor, learning rich visual features from a large dataset of unlabeled rice images. The downstream task consists of two branches: one for detecting the number of rice panicles and another for segmenting leaf areas. By combining these two branches, we are able to accurately assess the photosynthetic potential and reproductive capacity of rice plants. Experimental results demonstrate that our model outperforms traditional methods in both panicle detection and leaf area segmentation, achieving higher accuracy and robustness. We conduct experiments on a newly curated dataset, RiceVar, which comprises over 50,000 images covering three rice cultivars captured under varied angles and backgrounds. Our proposed method achieves a mean average precision of 81.401 in panicle detection, a 14.7 point improvement over ResNet50, and a Dice coefficient of 84.322 and intersection over union of 82.186 in leaf segmentation, outperforming the EAPT model by 14.08 and 2.45 points, respectively. Moreover, our model remains stable under varying environmental conditions, highlighting its practical value for rice breeding applications. By precisely evaluating panicle count and leaf width, our model supports the selection of high-yield, high-efficiency rice varieties, contributing to the advancement of sustainable agricultural practices. The relevant code and data are available at https://github.com/xiaobeial/Semi-supervised-detection-and-segmentation-algorithm-for-efficient-rice-breeding.