LGD-YOLO: An Asymmetric Lightweight Network with Dynamic Feature Alignment for Greenhouse Tomato Maturity Detection

Abstract

Greenhouse tomato detection faces critical challenges due to dense fruit occlusion, background interference, and the stringent computational constraints inherent to edge-deployed harvesting robots. Resolving these bottlenecks requires efficient architectures. We propose LGD-YOLO as an asymmetric lightweight network adapted from YOLOv10n specifically for edge-based tomato maturity detection. The architecture integrates a C2f-GMKSF module utilizing grouped multi-kernel convolutions to extract multi-scale textures with limited computational overhead. Precise feature alignment under occluded conditions is subsequently achieved through a Dy-HSFPN structure, accompanied by a C2f-CFCGLU module that expands the receptive field while preserving linear complexity. Furthermore, replacing the traditional detection head with a partial convolution head reduces memory access costs. A Focaler-Wise-SIoU loss function is utilized to stabilize bounding box regression against the lightweight penalty without introducing inference latency. Performance evaluations on a custom three-class dataset with a 180-image test set yield an 88.0% mAP@50. Relative to the baseline model, LGD-YOLO improves detection accuracy by 0.6 percentage points while shrinking the parameter volume by 37.6% to 1.41 M and lowering computational demand by 41.5% to 3.8 GFLOPs. Hardware deployment on an NVIDIA Jetson AGX Orin achieves a sustained processing speed of 40.6 FPS, while the weight file is 2.99 MB, supporting its feasibility for real-time agricultural robotics.