A lightweight detection network for immature green citrus via frequency domain sparsity and parameter sharing

Abstract

In unstructured orchard environments, detecting immature green citrus faces challenges such as background interference of similar colors, severe occlusion by branches and leaves, and dense distributions of small targets. These factors lead to low detection accuracy and poor performance in real time. To address these issues, a green citrus dataset under complex scenarios was constructed, and a lightweight detection model named LFD-YOLO is proposed in this paper. First, a dataset with semiautomatic annotation for training was efficiently constructed by integrating the Grounding DINO and SAM 2 visual foundation models. Based on this, improvements were made using YOLO11n as the baseline. Specifically, the C3k2-LFD module based on frequency dynamic convolution was designed to enhance fruit texture features under conditions of low contrast by utilizing the global receptive field and dynamic weighting in the frequency domain. Furthermore, the DySample upsampling operator was introduced, replacing the traditional nearest neighbor interpolation algorithm with a content-aware point sampling strategy to reduce computational redundancy while preserving the details of tiny targets. Finally, the lightweight residual shared detection head was constructed, utilizing parameter sharing and residual protection mechanisms to enhance the perception of occluded targets and significantly compress the model size. Experimental results demonstrate that LFD-YOLO achieves an mAP@0.5 of 91.14% and an F1 score of 0.863 on the newly constructed dataset, outperforming mainstream lightweight models such as YOLOv5n, YOLOv8n, YOLO11n, and YOLO12n. Compared with the baseline, LFD-YOLO increases the mAP@0.5 by 1.13 percentage points while simultaneously reducing the number of parameters and floating point operations by 19.8% and 23.8%, respectively. This method effectively achieves a balance between a lightweight architecture and detection accuracy, providing technical support for the edge deployment of agricultural harvesting robots.