A comparative study between deep learning approaches for aphid classification

Abstract

This study presents a performance comparison between two convolutional neural networks in the task of detecting aphids in digital images: AphidCV, customized for counting, classifying, and measuring aphids, and YOLOv8, state-of-the-art in real-time object detection. Our work considered 48,000 images for training for six different insect species (8,000 images divided into four classes), in addition to data augmentation techniques. For comparative purposes, we considered evaluation metrics available to both architectures (Accuracy, Precision, Recall, and F1-Score) and additional metrics (ROC Curve and PR AUC for AphidCV; mAP50 and mAP50-95 for YOLOv8). The results revealed an average F1-Score=0.891 for the AphidCV architecture, version 3.0, and an average F1-Score=0.882 for the YOLOv8, medium version, demonstrating the effectiveness of both architectures for training aphid detection models. Overall, AphidCV performed slightly better for the majority of metrics and species in the study, serving its design purpose very well. YOLOv8 proved to be faster to converge the models, with the potential to apply in research considering many aphid species.