Light-Inflorescence-Detection: A lightweight method for apple inflorescence detection in natural environments

Flower thinning effectiveness varies greatly across different phenological stages of apple inflorescences. Accurate detection of apple inflorescences and determination of their phenological stages can provide essential guidance for apple flower thinning. In this research, a method for detecting apple inflorescences at different stages based on the Light-Inflorescence-Detection (LID) object detection network was proposed. The Hybrid Group Block (HGBlock) was integrated into the Backbone, where the Parallel Convolution helped maintain feature diversity while effectively reducing the model’s computation and parameters. The Group Shuffle Cross Stage Partical-Neck (GSCSP-Neck) module was introduced in the Neck to achieve efficient feature extraction and model lightweighting through Group Convolution, Channel Shuffle, and Cross Stage Partical (CSP) connections. Shared Convolution was applied in the Head to achieve parameter reduction without compromising its accuracy. Evaluation on the apple inflorescence test set showed that the LID model reached an mean Average Precision (mAP) of 92.1%. It demonstrated superior efficiency over YOLOv7, YOLOv8n, YOLOv9, YOLOv10n, YOLOv11n, and RT-DETR, requiring only 3.32 MB of storage, 4.9 Giga Floating Point Operations (GFLOPs), and 1.56 M parameters. The detection speed on the NVIDIA Jetson AGX Orin Developer Kit was 38.10 Frames Per Seconds (FPS). These results demonstrated that the LID can perform accurate and fast detection of apple inflorescences across different phenological stages. This research serves as a reference for achieving efficient and accurate apple flower thinning.