Advanced seed classification and defect detection using convolutional neural networks and unsupervised learning techniques

Abstract

This paper introduces a hybrid deep learning framework for automated seed classification and defect detection in agricultural quality control. The system integrates supervised learning through Custom Convolutional Neural Networks (CustomCNN) and Attention-based CNNs (AttentionCNN) with unsupervised techniques including clustering algorithms and autoencoders. Working with high-resolution microscopic soybean seed images, the supervised models achieve test accuracies of 97.65% and 97.55% respectively for multi-class classification. The unsupervised autoencoder successfully identifies 312 anomalies through reconstruction error analysis. Clustering techniques partition RGB pixel data into three quality classes with silhouette scores up to 0.6810. This hybrid approach addresses key challenges in agricultural image analysis by enabling both accurate classification of labeled data and anomaly detection in unlabeled datasets, with processing times of 80-120ms per seed suitable for real-time deployment in seed sorting facilities.