Comparative analysis and evaluation of YOLO generations for banana bunch detection

Abstract

This study focuses on improving the automation of banana harvesting decisions for farmers with artificial intelligence assistance. Traditionally, experienced harvesters manually inspect fields to determine the optimal harvesting time, a process that is both labor-intensive and increasingly unsustainable due to a shortage of skilled workers. To address this challenge, this work proposes a computer vision-based approach for detecting banana bunches in images captured by mobile phones, as a preliminary step towards a comprehensive harvesting decision pipeline. To achieve this, a dataset was collected with 2179 photos of multiple Cavendish banana bunches in different light and exposure conditions, and a comparative analysis of You Only Look Once (YOLO) object detection models was conducted, from version 1 to 12, to identify the most accurate and efficient solution for banana bunch detection, ensuring compatibility with mobile-based applications. Among all models evaluated, YOLOv12n achieved the most balanced performance on five-fold cross-validation, with 93 % Average Precision (AP^50test), 51 % AP^50–95test, and 5.1 ms latency, making it well-suited for real-time deployment on resource-constrained edge devices.