Analysis of YOLO26 Model Performance with Transfer Learning in Detecting Coffee Bean Defects

Abstract

Background: Indonesia is one of the world's largest coffee producers, yet post-roasting coffee bean defects remain a critical challenge that reduces product quality and market competitiveness. Manual sorting processes are inconsistent and prone to human visual limitations. Objective: This study aims to analyze the performance of the YOLO26 nano model with transfer learning for detecting and classifying post-roasting coffee bean defects, and to evaluate the effect of grid search-based hyperparameter tuning on model performance. Methods: A first-party dataset of 4,567 images covering five defect categories — insect damage, under roast, quaker, nugget, and shell — and one non-defect category was collected from three local Indonesian coffee roasting companies. After augmentation, the dataset expanded to 10,595 images with a 70:20:10 training-validation-testing split ratio. YOLO26, a deep learning-based object detection model released in January 2026, was applied using transfer learning and optimized through grid search hyperparameter tuning across optimizer, learning rate, epoch, classification loss, and weight decay configurations. Results: The model was evaluated using precision, recall, F1 score, mean Average Precision at IoU 50 (mAP50), and mean Average Precision at IoU 50-95 (mAP50-95) on the test dataset, with results demonstrating competitive multi-class detection performance across all defect categories. Conclusion: YOLO26 nano with transfer learning and hyperparameter tuning is a viable approach for automated post-roasting coffee bean defect detection, contributing to quality control advancements in the Indonesian coffee industry.