YOLO-RDM: Innovative Detection Methods for Eggplants and Stems in Complex Natural Environment

Accurate identification of eggplants and their stems in complex environments is crucial for the research on intelligent harvesting equipment for eggplants. Current object detection methods have limited and traditional approaches to eggplant research. To achieve rapid identification of eggplants and stems in complex environments and enhance overall accuracy, this study establishes a database for eggplants and stems and proposes an efficient detection model based on the YOLOv8n network. First, a lightweight Receptive-Field Attention Convolution (RFAConv) and Mixed Local Channel Attention (MLCA) attention mechanism are used to design the C2f_RM module, replacing the C2f module in YOLOv8n to create a lightweight yet effective model. Next, the DD-Head structure reparameterization detection head module is integrated, effectively merging deep and shallow information while retaining small-scale target feature information to improve detection accuracy. Finally, the loss function is substituted with the MPDIoU loss, optimizing the bounding box regression algorithm and enhancing convergence speed and regression accuracy. The proposed YOLO-RDM model (C2f_RM, DD-Head, MPDIoU) attains an average accuracy of 93.6% on the dataset, representing a 3.4% improvement over the original YOLOv8n model. Detection accuracies for eggplants and stems increased by 1.1% and 5.8%, respectively. Additionally, the F1 score improved by 4.96%, reaching 89.09%. This model demonstrates superior detection performance and robustness for both eggplants and their stems, providing valuable technical support for the automation of eggplant harvesting.