Authors

Abstract

Fruit instance segmentation algorithms are critical for target localization in fruit-picking robots, enabling accurate area estimation of fruits within images. However, existing methods often face limitations such as high computational cost, poor adaptability to low-power devices, and reduced detection performance in complex environments. To address these challenges, YOLOSeg-Jujube was proposed - an improved instance segmentation algorithm based on YOLOv8 - validated on a jujube image dataset annotated with bounding polygons. The architecture of YOLOSeg-Jujube was optimized through ablation studies to identify the most effective structure. The final network design includes Focus, CBS, Conv4cat, SPD, and SPPFr modules as the backbone, and a YOLOv8 head incorporating the SIoU loss function. Compared to YOLOv4-tiny, YOLOv5n, YOLOv7-tiny, and YOLOv8n, YOLOSeg-Jujube achieves reductions in parameter size of 62.8%, 9.8%, 73.9%, and 23.6%, respectively. The model achieves 83.5% B_mAP and 83.2% S_mAP, outperforming mainstream YOLO variants in both segmentation accuracy and area estimation of target objects. YOLOSeg-Jujube is robust, fast, and computationally efficient, making it suitable for deployment on resource-constrained platforms. Furthermore, it demonstrates strong potential for recognizing ripeness stages of fresh jujubes, providing technical support for intelligent harvesting systems in the field.

Abstract in Chinese

水果实例分割算法是水果采摘机器人目标定位的重要方法。能够确保对图像中的目标进行正确的区域估算。与此同时，模型在复杂农业环境下计算成本高、不便于在低功耗计算设备上部署等都对水果采摘机器人用实例分割造成限制。因此，为了解决这些问题，本研究在 YOLOv8 的基础上设计了 YOLOseg-Jujube，并在自然光照环境下的红枣图像数据集上进行了验证。YOLOseg-Jujube 网络由 Focus、CBS、Conv4cat、SPD 和 SPPFr 组成骨干网络，YOLOv8 head 和 SIoU loss 组成头部网络。通过对比，YOLOseg-Jujube 的参数分别比 YOLOv4-tiny、YOLOv5n、YOLOv7-tiny 和 YOLOv8n 低 62.8%、9.8%、73.9% 和 23.6%。YOLOseg-Jujube 的 B_mAP 为 83.5%，S_mAP 为 83.2%，YOLOseg-Jujube 在分割目标的面积估计方面性能优于 YOLO 主流变体的算法。YOLOseg-Jujube 算法稳健、快速、准确，计算成本低，并且能够识别枣果的成熟阶段，可为田间鲜枣采摘机器人的研究提供技术支撑。