Scholarlink

<p>The quality evaluation of processed rice grains is an important factor in determining market acceptance, pricing, storage stability, processing quality, and overall consumer approval. In the literature, the machine vision methods are predominantly based on supervised machine learning which rely on manual labelling, and therefore, face issues such as time intensiveness, subjectiveness and overlapping classes. Therefore, in this study, a deep unsupervised method Contrastive-RC is developed that leverages contrastive self-supervised learning technique for fine-grained damage classification of milled white rice. Particularly, contrastive-RC is structured to combine SimCLR, UMAP, and HDBSCAN, where self-supervised contrastive learning (SimCLR) is used for feature representation followed by dimensionality reduction using UMAP, and clustering using HDBSCAN. To enable this, a dataset of 20,102 high magnification images (24 MP at 3.8&#xa0;μm/pixel) of individual rice grains spread across six different types of surface damages is developed. It is demonstrated that the Contrastive-RC successfully clusters the rice grains into six main classes with well-defined attributes, namely normal<i>-damage</i>, <i>chalky-discoloured</i>, <i>discoloured</i>,<i> half-chalky</i>, <i>healthy</i>, and <i>broken</i>. The contrastive-RC achieves this clustering with an accuracy of 0.88, macro-F1 score of 0.82 and a silhouette score of 0.599, indicating a high clustering effectiveness in terms of clear separation between the clusters and their purity. It is demonstrated that the contrastive-RC method can be extended to subclassify these damages based on the damage severity by providing a low-level control, thereby enabling the method to be used in multiple use-cases. The method is fast, versatile and robust towards changes in variables like brightness and grain orientation, thus making it ideal for real world use and extension to other varieties of milled rice. Finally, a comparison of the unsupervised contrastive-RC method with both the existing methods such as K-means and t-SNE, and the supervised CNN-approach is presented. It is shown that the contrastive-RC method outperforms K-means across all the performance metrics (i.e., accuracy: 88% vs. 78%, macro-F1 score: 0.82 vs. 0.67, ARI, NMI and silhouette score: 0.599 vs. 0.523). Further in comparing with CNN-based supervised methods, the contrastive-RC method performs better ins terms of the ability to handle 2.5 times greater number of images, handle data set imbalance due to clustering resilience, and offering a nominally high accuracy of 88% (compared to 98% with CNN) with an unlabelled data.</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

The characterization of damages and their severity in milled rice by applying unsupervised learning to a high-magnification image dataset

  • 期刊论文

  • 发表日期：

    2025年8月8日

查看全文
  • Kriz Moses,
  • Isprash Chauhan,
  • Bhupendra,
  • Hitarth Kankar,
  • Ankur Miglani

摘要

The quality evaluation of processed rice grains is an important factor in determining market acceptance, pricing, storage stability, processing quality, and overall consumer approval. In the literature, the machine vision methods are predominantly based on supervised machine learning which rely on manual labelling, and therefore, face issues such as time intensiveness, subjectiveness and overlapping classes. Therefore, in this study, a deep unsupervised method Contrastive-RC is developed that leverages contrastive self-supervised learning technique for fine-grained damage classification of milled white rice. Particularly, contrastive-RC is structured to combine SimCLR, UMAP, and HDBSCAN, where self-supervised contrastive learning (SimCLR) is used for feature representation followed by dimensionality reduction using UMAP, and clustering using HDBSCAN. To enable this, a dataset of 20,102 high magnification images (24 MP at 3.8 μm/pixel) of individual rice grains spread across six different types of surface damages is developed. It is demonstrated that the Contrastive-RC successfully clusters the rice grains into six main classes with well-defined attributes, namely normal-damage, chalky-discoloured, discoloured, half-chalky, healthy, and broken. The contrastive-RC achieves this clustering with an accuracy of 0.88, macro-F1 score of 0.82 and a silhouette score of 0.599, indicating a high clustering effectiveness in terms of clear separation between the clusters and their purity. It is demonstrated that the contrastive-RC method can be extended to subclassify these damages based on the damage severity by providing a low-level control, thereby enabling the method to be used in multiple use-cases. The method is fast, versatile and robust towards changes in variables like brightness and grain orientation, thus making it ideal for real world use and extension to other varieties of milled rice. Finally, a comparison of the unsupervised contrastive-RC method with both the existing methods such as K-means and t-SNE, and the supervised CNN-approach is presented. It is shown that the contrastive-RC method outperforms K-means across all the performance metrics (i.e., accuracy: 88% vs. 78%, macro-F1 score: 0.82 vs. 0.67, ARI, NMI and silhouette score: 0.599 vs. 0.523). Further in comparing with CNN-based supervised methods, the contrastive-RC method performs better ins terms of the ability to handle 2.5 times greater number of images, handle data set imbalance due to clustering resilience, and offering a nominally high accuracy of 88% (compared to 98% with CNN) with an unlabelled data.