LemonDet: A Lightweight YOLO11 Architecture with Dataset-Specific Color Priors for Small Lemon Detection in Orchards

Accurate and reliable fruit detection in agricultural fields is crucial for applications such as yield estimation, crop tracking, and autonomous harvesting. However, the high density of small fruits in images, overlap, and complex vegetation can limit object detection performance. In this study, LemonDet, a parameter-efficient object detection model based on YOLO11n, is proposed to improve the detection of small lemons in particular. Considering the object scale distribution in the dataset, the standard P3-P5 detection structure is restructured as P2-P4 to ensure the preservation of high-resolution spatial features. In addition, the Lemon Color Prior Convolution (LCP-Conv) module, which transfers the dataset-specific RGB color prior obtained from labeled lemon regions in the training data to the early feature extraction process, has been developed. A label-guided local image enhancement approach, applied only to training images, has also been included in the model to strengthen the limited pixel representations of small objects. In the experimental results, LemonDet achieved 87.2% Precision, 81.6% Recall, 84.3% F1-score, 89.2% mAP50, and 54.8% mAP50-95. With 0.99 million parameters, the model provides a more parameter-efficient architecture than the baseline YOLO models while achieving higher detection performance. Ablation results show that color normalization and label-guided local enhancement contribute to performance. The findings indicate that considering object scale distribution and dataset-specific color information together in model design is an effective approach for detecting small lemons.