Tomato detection in challenging scenarios using YOLO-based single stage detectors

Abstract

Fruit detection is a fundamental task for several agricultural applications, including robotic harvesting, yield estimation, and precision farming in general. However, uneven field conditions, such as branch and leaf occlusions, lighting variations, and shading, make this task particularly challenging. This paper proposes a deep learning approach to automatically segment tomato berries, using transfer learning from two state-of-the-art deep neural networks, namely YOLOv5 and YOLOv8, previously trained on images acquired under controlled conditions. The results obtained for images captured under various field conditions by a consumer-grade camera demonstrate that the models achieve relatively high classifying performance despite the low quality of the input images and the challenges arising from reduced object size, similarity between objects, and their color, fruit cluster occlusions, and highly variable illumination conditions.