Abstract: Accurate identification of seedless white grape stem under small trellises is the key to the automatic picking task of grape picking robot. Aiming at the problem of poor recognition effect of fresh seedless white grape stem under Small trellises in Turfan City, Xinjiang, a seedless white grape stems recognition model Small—YOLO based on YOLOv8 was proposed to realize automatic recognition of seedless white grape stem, improve the target detection head on the original model structure, and improve the feature fusion of shallow and deep networks. The ability of extracting information from the seedless white grape stem was increased. Deformable convolutional DCN is used in the shallow network to enhance the modeling ability of convolution operation on morphological changes, so that the convolutional kernel can better adapt to the irregular deformation in the input feature map, which is helpful to improve the performance of the visual model when dealing with targets of different scales, shapes and transformations, increase the coordinate attention mechanism CA, optimize the accuracy rate of the seedless white grape peduncle recognition. The test results showed that the mAP value of the improved recognition model for the seedless white grape stem reached 76.2%. Compared with YOLOv3—tiny, YOLOv5n, YOLOv6, YOLOv7, YOLOv8n and other algorithms, the mAP value was improved by 23.9%, 8%, 7.6%, 9.2%, 7%, respectively, while maintaining a fast detection speed, realizing the possibility of the seedless white grape mechanical picking under small tallows.

Key words: seedless white grapes, harvesting robot, stem recognition, coordinate attention mechanism, deformable convolution, visual model

摘要： 小棚架下准确识别无核白葡萄果梗是葡萄采摘机器人自动采摘任务的关键。针对新疆吐鲁番市小棚架下鲜食无核白葡萄果梗识别效果差的问题，提出一种基于YOLOv8的无核白葡萄果梗识别模型Small—YOLO，实现无核白葡萄果梗的自动识别。在原有的模型结构上改进目标检测头，提高浅层与深层网络的特征融合，增加对无核白葡萄果梗信息提取的能力。在浅层网络中采用可变形卷积DCN以增强卷积操作对形态变化的建模能力，使得卷积核可以更好地适应输入特征图中的不规则变形，有助于提高视觉模型在处理不同尺度、形态和变换目标时的性能。增加坐标注意力机制CA，优化无核白葡萄果梗识别的准确率。结果表明，改进后的识别模型对无核白葡萄果梗平均精度mAP值达到76.2%。与YOLOv3—tiny、YOLOv5n、YOLOv6、YOLOv7、YOLOv8n等算法相比，mAP值分别提升23.9%、8%、7.6%、9.2%、7%，同时保持较快的检测速度，实现在小棚架下无核白葡萄机械采摘可能性。

关键词: 无核白葡萄, 采摘机器人, 果梗识别, 坐标注意力机制, 可变形卷积, 视觉模型