Design and Development of A Sorting System for Ripeness and Damage in Avocado (Persea Americana Mill.) Based on Machine Learning

Abstract

Avocado is an important commodity that requires proper postharvest handling, particularly in
sorting based on ripeness levels and fruit condition. Conventional methods that still rely on
manual inspection have limitations in terms of accuracy and efficiency. This study proposes a
real-time sorting system based on YOLOv8n running on a Raspberry Pi 4 Model B with a
Google Coral accelerator to detect avocados and classify their ripeness levels as well as damage.
The model was trained with a custom dataset and achieved a precision of 77.2%, recall of
84.4%, and a macro-average F1-Score of 80.6%. The system was integrated with a camera for
image acquisition and a mini conveyor for the sorting process. Experimental results
demonstrated reliable detection with an inference speed of 11.7 ms per image, and field testing
successfully classified avocados according to their categories. This system has proven to be
effective, cost-efficient, and supports the improvement of avocado postharvest quality.