A Database for Soybean Seed Classification

This paper proposes a novel public database for soybean seed defect classification. The database is publicly available and features seven defect classes labeled based on visual characteristics by domain specialists. Each seed sample includes three images taken from random rotations to ensure a comprehensive representation. Additionally, seeds from three different locations, each labeled by different experts, are included in the database. No public currently available soybean defect classification dataset provides location information. This allows for the evaluation of the generalization capabilities of the model on seeds collected on previously unseen regions. Experiments were conducted and compared with the GBl352 database to validate our approach. We employed handcrafted texture descriptors and non-handcrafted features extracted from established convolutional neural network architectures, using the network's last layer activations as the feature set. The results show that partitioning folds by the regions where seeds were collected significantly impact classification performance. Employing the SVM classifier and the product prediction fusion rule, we achieved an F1-score of 85.06% with region-based cross-validation folds and 95.89% without region-based fold partitioning. In experiments involving two classes (intact vs. defective), we achieved an F1-score of 99.32% with region-based cross-validation folds. We expect this database to foster the development of more robust models capable of generalizing to previously unseen regions.