Fruit detection with yolo algorithm

In the agricultural sector, the use of artificial intelligence and deep learning-based image processing techniques is becoming increasingly widespread in order to enhance productivity and reduce dependency on human labor. Particularly in applications such as fruit detection and harvest automation, object detection algorithms play a critical role. In this thesis study, various versions of the You Only Look Once (YOLO) algorithm (YOLOv8, YOLOv9, YOLOv10, YOLOv11) are comparatively analyzed with the aim of accurately detecting kiwifruit in complex agricultural environments. Considering environmental challenges such as dense foliage, shading, overlapping fruit clusters, and varying lighting conditions—where traditional image processing techniques often fall short—YOLO-based models were employed as a solution. The models were trained using both default configurations and optimized hyperparameters, and their performance was evaluated based on accuracy and inference stability criteria. A dataset consisting of 40 naturally captured images and 1,936 labels, sourced from the literature, was used for testing. Among the developed models, YOLOv11s achieved the highest performance with 95.6% mAP@0.5, an F1 score of 89.1%, recall of 89.9%, and precision of 88.3%. Real-time performance tests were conducted on the NVIDIA Jetson TX2 platform, where YOLOv11s demonstrated an average inference time of 3.33 seconds, making it a suitable and field-deployable solution compatible with embedded systems. In conclusion, YOLO-based models have emerged as the most effective solutions for kiwifruit detection in terms of both accuracy and operational efficiency. This study contributes significantly to the development of fast and reliable AI-supported object detection systems in agricultural production environments, offering practical and effective methods.