Research on an Enhanced YOLOV8-Seg Model for the Lilac Harvesting Localization Algorithm

Lilacs not only have ornamental value but also extensive application value. Therefore, how to achieve nondestructive picking of lilacs is of great significance. Identifying the correct picking location is crucial for achieving non-destructive picking, among which accurately locating the lilac flowers and their stems is the core of determining the picking point. Therefore, in view of the picking requirements for intact lilacs and the picking location problem of lilacs in complex environments, this paper proposes a picking location algorithm based on the improved YOLOV8-Seg model. Aiming at the occlusion problem of lilac flowers in complex environments and the problem of small stem detail features, the learning effect is improved by optimizing the hyperparameters, the ParNetattention attention mechanism is introduced to enhance the ability of extracting key features, and the LightConv module is used for lightweight processing. The ablation experiments indicated that the average accuracies of Box(MAP50), Box(F1), Mask(MAP50), and Mask(F1) for flowers and stems increased to $87.5 \%, 0.83,88.3 \%$, and 0.84 respectively; Among them, the segmentation accuracies of lilac stems, $\mathrm{Bj}(\mathrm{AP} 50)$, $\mathbf{B j}(\mathbf{F} 1), \mathbf{M j}(\mathbf{M A P 5 0})$, and $\mathbf{M j}(\mathbf{F} 1)$, increased by $33.9 \%, 20.97 \%$, $33.9 \%$, and $26.67 \%$ respectively.