Comparison of ResNet and YOLO Models for Computer Vision-Based Grading of Chilean Apples for Agroexport

Sanchez G.N.V.; Bailon J.A.M.; Espinoza C.H.I.; Aza R.F.A. · 2026 The 9th International Conference on Image and Graphics Processing, ICIGP 2026 (2026) · DOI: 10.1109/icigp68997.2026.11619820

Resumen

This paper presents a rigorous comparative analysis of computer vision models with the objective of identifying the most suitable architecture for the postharvest classification of Chilean apples destined for agroexport. Four architectures, ResNet-50, YOLOv5, YOLOv8 and YOLOv11, were evaluated on two tasks: estimating surface color, percentage of red area, and detecting visible defects, bruises or cuts. A dataset of 605 labeled images covering three commercial categories, Extra Premium, Apple G1 and Defective, was used with a stratified train- validation-test split. YOLOv5 achieved the best performance, with 95.8 percent overall accuracy and 94 percent mean average precision at 0.5, and an average processing time of 510.73 milliseconds per apple. These results indicate that modern detection models, particularly YOLOv5, provide robust performance for automatic postharvest classification systems in agroexport settings.

Instituciones: Universidad Peruana de Ciencias Aplicadas