Abstract The internal structure of a seed plays a vital role during germination. Hence, before planting the seed internal quality inspection is advantageous to produce healthy seedlings. In this study, an X-ray CT scanner was used to generate CT images of five-year-old naturally aged pepper seeds. Several processing techniques such as reslicing, feature extraction, and classification were performed on these images. The reslicing process was applied to construct three different planes, namely, transaxial, sagittal, and coronal plane images from raw CT images. Then, three images were selected from each sample (one from each plane) for feature extraction. Using a pattern recognition algorithm, the gray-level co-occurrence matrix (GLCM) was created for each image, and twenty-two types of statistical derivations were performed to generate GLCM textural features. A supervised data matrix was constructed based on the germination results of the seed samples from the images, where the seeds were divided into two classes: normal viable seeds (class-1) and nonviable & abnormal viable seeds (class-2). Supervised classification methods, such as partial least-squares discriminant analysis (PLS-DA), support vector machine (SVM), and K-nearest neighbor (KNN) were used to evaluate the best outcome. Among the tested classifiers, PLS-DA provided the highest accuracy of 88.7% with five-fold cross validation, where seven important features extracted from different angles (θ) were found from the beta coefficient to be significant in the classification of the seed. Results from this study show that X-ray CT imaging incorporated with a pattern recognition system is a robust technique to classify seeds based on their internal quality attributes.

中文翻译：

---

X射线CT成像基于内部结构的辣椒种子品质分类

摘要 种子的内部结构在萌发过程中起着至关重要的作用。因此，在种植前进行种子内部质量检查有利于生产健康的幼苗。在这项研究中，使用 X 射线 CT 扫描仪生成 5 年自然老化辣椒种子的 CT 图像。对这些图像执行了多种处理技术，例如重新切片、特征提取和分类。应用重新切片过程从原始 CT 图像构建三个不同的平面，即经轴、矢状和冠状平面图像。然后，从每个样本中选择三张图像（每个平面一张）进行特征提取。使用模式识别算法，为每个图像创建灰度共生矩阵（GLCM），并进行了 22 种类型的统计推导以生成 GLCM 纹理特征。基于图像种子样本的萌发结果构建监督数据矩阵，其中种子分为两类：正常活种子（类1）和非活和异常活种子（类2）。使用监督分类方法，例如偏最小二乘判别分析 (PLS-DA)、支持向量机 (SVM) 和 K-最近邻 (KNN) 来评估最佳结果。在经过测试的分类器中，PLS-DA 通过五重交叉验证提供了 88.7% 的最高准确率，其中从 beta 系数中发现从不同角度 (θ) 提取的七个重要特征在种子分类中是显着的。