FLD-YOLO: A lightweight and efficient framework for detecting wolfberry fruits and pedicels in unstructured agricultural environments

Manual picking of wolfberries is labor-intensive and inefficient. Precise real-time detection is a key prerequisite for developing automated picking machinery. To address the challenges in unstructured agricultural environments, this paper proposes a lightweight and efficient detection framework based on FLD-YOLO (Feature-focusing, Lightweight, Deformable-YOLO). Wolfberry fruits and pedicels exhibit significant morphological differences and are easily occluded by branches and leaves, which makes real-time accurate detection difficult. In the framework, a deformable attention mechanism is employed to adjust the receptive field. A pyramid network is also used to enhance multi-scale fusion. Additionally, a lightweight detection head is combined with pruning to achieve model compression. These features reduce the number of parameters and improve detection efficiency. On a self-constructed wolfberry dataset of 5,760 images, experimental results show that FLD-YOLO achieves an mAP50 of 82.4% and an mAP50-95 of 55.5% on the test set. The inference time is only 2.2 ms per image. Compared to the baseline model, this framework improves mAP50 by 1.7 percentage points. It also reduces the parameter count by 43.8% (to 1.45 M). Field trial results indicate that FLD-YOLO maintains high fruit detection precision (AP50 > 90%). It significantly improves the missed detection issue for subtle pedicels. This elevates the pedicel detection precision (AP50 Stem) to 68.0%, which is markedly superior to RT-DETR (57.4%) and YOLOv3-tiny (53.2%). In unstructured agricultural environments, the model’s overall recognition accuracy reaches 87.5%. The research results demonstrate that this framework has strong adaptability to small targets. It effectively addresses issues of large morphological differences and occlusion by branches and leaves. It can be deployed on edge computing devices or integrated into real-time autonomous picking robots. It also provides technical support for the detection of other small fruit crops.