Scholarlink

<p>This study presents an accurate hybrid classification approach for dry bean varieties by integrating a lightweight one-dimensional Residual Network (1D-ResNet) with pre-extracted morphological features. The proposed architecture is designed to investigate how residual learning contributes to incremental feature refinement and stable optimization in low-dimensional morphological tabular data. By incorporating skip connections, the model facilitates effective modeling of nonlinear feature interactions while maintaining training stability. The framework operates directly on geometric and shape-based descriptors without relying on image-based representation learning. Experimental results over 10 independent runs demonstrate an average classification accuracy of 95.70%, outperforming traditional machine learning methods such as Support Vector Machine (93.13%) and XGBoost (93.00%), as well as deep learning baselines including DBANN2 (93.44%) and 1D-CNN (94.79%). These findings highlight the effectiveness of residual structures in enhancing interaction modeling and optimization behavior for morphology-driven tabular classification tasks.</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

Lightweight 1D-ResNet architecture for improving dry bean variety classification

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Thi-Thu-Hong Phan

摘要

This study presents an accurate hybrid classification approach for dry bean varieties by integrating a lightweight one-dimensional Residual Network (1D-ResNet) with pre-extracted morphological features. The proposed architecture is designed to investigate how residual learning contributes to incremental feature refinement and stable optimization in low-dimensional morphological tabular data. By incorporating skip connections, the model facilitates effective modeling of nonlinear feature interactions while maintaining training stability. The framework operates directly on geometric and shape-based descriptors without relying on image-based representation learning. Experimental results over 10 independent runs demonstrate an average classification accuracy of 95.70%, outperforming traditional machine learning methods such as Support Vector Machine (93.13%) and XGBoost (93.00%), as well as deep learning baselines including DBANN2 (93.44%) and 1D-CNN (94.79%). These findings highlight the effectiveness of residual structures in enhancing interaction modeling and optimization behavior for morphology-driven tabular classification tasks.