Grapevine-Seg: A grapevine segmentation method based on an improved YOLACT

Real-time segmentation of grapevine cordons and shoots remains a critical bottleneck for automated viticultural management, particularly in resource-constrained field environments where computational efficiency is paramount. This study established a lightweight segmentation model based on an improved YOLACT (You Only Look At Coefficient) framework, optimised for real-time grapevine segmentation on embedded systems. In place of the original ResNet backbone is Ghost-Net, which uses Ghost modules to generate redundant feature maps and produce richer feature representations, thus reducing the computational load of convolutions and improving small-object detection and segmentation. Concurrently, an EMAttention (Efficient Multi-scale Attention) module is embedded at the skip connections between the Feature Pyramid Network (FPN) and the Mask Head. This module concatenates the FPN features of multiple scales with Mask Head features, applies spatial convolution to generate spatial attention maps that bolster object-region features, and adaptively fuses multi-scale features with learned weights, thus improving segmentation performance across differently sized objects. The modifications, focusing on lightweight design and inference speed, yield a model that balances accuracy and efficiency more suitably for embedded viticultural systems than existing benchmarks. Under identical experimental conditions, the improved model outperformed mainstream segmentation frameworks. The average detection accuracies for grapevine and lateral shoot test samples reached 69.46 % and 67.66 %, respectively. These results demonstrate the potential of this approach for enabling real-time, field-deployable grapevine segmentation in precision viticulture applications. The annotated dataset (Grapevine-Seg) is publicly available at https://zenodo.org/records/18218165.