Object detection method for kiwifruit (Actinidia deliciosa) based on improved YOLO11x-mod

Kiwi is one of the most important agricultural products in Turkey, and its maturity directly determines the quality and market value of the product. Therefore, there is a need for an accurate and reliable detection system that can minimize post-harvest losses while increasing productivity. In this study, a dataset of 420 images taken under different environmental conditions was expanded to 928 images using data augmentation techniques, and a total of 22 750 kiwi samples were labeled. In the study, the latest deep learning architectures, YOLOv8, YOLOv10, and YOLO11, were systematically compared under identical conditions. As a result of these comparisons, the YOLO11x-mod model showed the highest performance when optimized with hyperparameter strategies such as AdamW optimization, low weight decay, appropriate learning rate, and dropout-mosaic augmentation techniques. This model achieved 81.76% accuracy, 83.97% recognition rate, 86.19% mAP@50, and 66.80% mAP@50:95, delivering excellent results, especially in challenging scenarios such as dense foliage, object overlap, and variable lighting conditions. In addition, the model’s operation with 72.5 million parameters, 272 GFLOPs, and 42 fps was found to be suitable for real-time applications, balancing accuracy and speed. The insights gained from the study establish the YOLO11x-mod model as a new reference point for kiwi detection and directly address limitations in the literature regarding insufficient detection of small or overlapping fruits. In this respect, the results of this study not only contribute to science but also show great potential for the integration of the model into robotic harvesting systems and precision agriculture applications.