Scholarlink

To address the inefficiency and limited accuracy of broken rice rate detection in rice processing, this study proposes an enhanced DenseNet121 model integrated with a Spatial-Channel Cooperative Attention (SACA) mechanism. The preprocessing pipeline is optimized using affine transformation and the watershed algorithm to improve image correction and segmentation. The SACA module, embedded into the DenseNet121 structure, combines multi-scale dilated convolutions and multi-head self-attention to enhance sensitivity to subtle discriminative features such as rice grain cracks. Experimental results show that the improved model achieves a classification accuracy of 99.14% on a custom-built dataset, 1.24% higher than the baseline DenseNet121, with parameters reduced to only 30.5% of ResNet50. By incorporating a projected area-mass density model, a quantitative broken rice rate analysis framework aligned with national standards is established, offering a robust solution for intelligent and precise quality control in grain processing.

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

CNN-Based Grain Quality and Quantity Prediction from Small Samples

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jiayuan Gao,
  • Zhen Zhou,
  • Yanbo Chen,
  • Lei Zhang,
  • Xiaoqiang Chen,
  • Bo Mao

摘要

To address the inefficiency and limited accuracy of broken rice rate detection in rice processing, this study proposes an enhanced DenseNet121 model integrated with a Spatial-Channel Cooperative Attention (SACA) mechanism. The preprocessing pipeline is optimized using affine transformation and the watershed algorithm to improve image correction and segmentation. The SACA module, embedded into the DenseNet121 structure, combines multi-scale dilated convolutions and multi-head self-attention to enhance sensitivity to subtle discriminative features such as rice grain cracks. Experimental results show that the improved model achieves a classification accuracy of 99.14% on a custom-built dataset, 1.24% higher than the baseline DenseNet121, with parameters reduced to only 30.5% of ResNet50. By incorporating a projected area-mass density model, a quantitative broken rice rate analysis framework aligned with national standards is established, offering a robust solution for intelligent and precise quality control in grain processing.