Scholarlink

Corn ear phenotypic characteristics are important predictors of corn quality, yield, and growth. The Corn Instance Segmentation Model (CSM) is a high-precision instance segmentation framework that was created in this study to meet the need for automated phenotypic extraction and crop monitoring. Using accurate localization supplied by an RT-DETR-based model to direct segmentation through a SAM network, CSM combines sophisticated object detection and instance segmentation techniques. In order to maximize multi-scale feature fusion, the framework includes context and spatial feature calibration modules (CFC and SFC) in addition to a dilated residual network (DWRP3) for feature extraction. LoRA fine-tuning is also used to increase model adaptability and computational efficiency. This method improves operating efficiency by drastically lowering manual involvement while simultaneously increasing segmentation accuracy. According to experimental results on a custom corn ear dataset, CSM outperforms Mask R-CNN, Mask2Former, and YOLACT in precision by 3.5%, 4.3%, and 12.7%, respectively, achieving 95.5% precision, 84.8% mean intersection-over-union (mIoU), and 88.4% recall. With great potential for widespread use, this research offers crucial technological assistance for automated crop monitoring in precision agriculture.

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

CSM: Corn Instance Segmentation Model Fusing Dilated Residual Networks and Low-Rank Adaptation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Deli Zhu,
  • Xinjie Li,
  • Liang He

摘要

Corn ear phenotypic characteristics are important predictors of corn quality, yield, and growth. The Corn Instance Segmentation Model (CSM) is a high-precision instance segmentation framework that was created in this study to meet the need for automated phenotypic extraction and crop monitoring. Using accurate localization supplied by an RT-DETR-based model to direct segmentation through a SAM network, CSM combines sophisticated object detection and instance segmentation techniques. In order to maximize multi-scale feature fusion, the framework includes context and spatial feature calibration modules (CFC and SFC) in addition to a dilated residual network (DWRP3) for feature extraction. LoRA fine-tuning is also used to increase model adaptability and computational efficiency. This method improves operating efficiency by drastically lowering manual involvement while simultaneously increasing segmentation accuracy. According to experimental results on a custom corn ear dataset, CSM outperforms Mask R-CNN, Mask2Former, and YOLACT in precision by 3.5%, 4.3%, and 12.7%, respectively, achieving 95.5% precision, 84.8% mean intersection-over-union (mIoU), and 88.4% recall. With great potential for widespread use, this research offers crucial technological assistance for automated crop monitoring in precision agriculture.