Scholarlink

<p>Rapid and accurate rice segmentation is essential for automated phenotypic analysis in quality inspection; however, existing methods often suffer from high computational complexity or limited robustness to adherent grains, hindering real-time industrial deployment. To address this issue, we propose a lightweight instance segmentation model, Patch in Batch and YOLACT (PiB-Yolact), which integrates target region matching and local weight sharing for efficient rice quality detection. Region matching reduces frame redundancy and improves detection continuity, while local weight sharing enhances feature representation for accurate segmentation of adherent grains. Rice image datasets covering four rice types-short grain rice, waxy rice, long grain rice, and bulk rice-were constructed, with 500 manually annotated images per category augmented to 3,000 images. Experimental results show that PiB-Yolact achieves a segmentation precision of 87.06% at 4.10 frames per second, outperforming the standard YOLACT model (83.66% at 4.87 frames per second) in accuracy while preserving real-time capability. The model estimates key quality indicators, including grain count, chalkiness rate, and broken rice rate, with errors below 1% compared to ground-truth data. Future work will focus on further improving segmentation accuracy and efficiency under challenging conditions, as well as extending the model to support additional rice quality indicators.</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

A rapid segmentation method for rice quality inspection via PiB-Yolact

  • 期刊论文

  • 发表日期：

    2026年2月21日

查看全文
  • Xiangjun Duan,
  • Zhongyin Sheng,
  • Shan Zeng,
  • Bing Li,
  • Jinbo Zhou,
  • Weiqiang Yang

摘要

Rapid and accurate rice segmentation is essential for automated phenotypic analysis in quality inspection; however, existing methods often suffer from high computational complexity or limited robustness to adherent grains, hindering real-time industrial deployment. To address this issue, we propose a lightweight instance segmentation model, Patch in Batch and YOLACT (PiB-Yolact), which integrates target region matching and local weight sharing for efficient rice quality detection. Region matching reduces frame redundancy and improves detection continuity, while local weight sharing enhances feature representation for accurate segmentation of adherent grains. Rice image datasets covering four rice types-short grain rice, waxy rice, long grain rice, and bulk rice-were constructed, with 500 manually annotated images per category augmented to 3,000 images. Experimental results show that PiB-Yolact achieves a segmentation precision of 87.06% at 4.10 frames per second, outperforming the standard YOLACT model (83.66% at 4.87 frames per second) in accuracy while preserving real-time capability. The model estimates key quality indicators, including grain count, chalkiness rate, and broken rice rate, with errors below 1% compared to ground-truth data. Future work will focus on further improving segmentation accuracy and efficiency under challenging conditions, as well as extending the model to support additional rice quality indicators.