Abstract: In order to realize the intelligent picking of Myrica rubra, a method based on the improved YOLOX-NANO algorithm is proposed to detect the ripeness of Myrica rubra. By introducing the ECA channel attention module into the featureenhanced extraction network, the networks ability to extract channel features is enhanced. The Focus loss function is introduced to replace the standard crossentropy loss function, which solves the problem of unbalance ofthe positive and negative samples in the singlestage network, and avoids the gradient direction pointing to a nonoptimal solution. The EIoU loss function is used to improve the accuracy of the network model for target recognition. The experimental results show thatcompared with the original YOLOX-NANO, the improved YOLOX-NANO algorithm has improved the recognition accuracy for three different ripeness of prune fruits with mAP of 92.67%. and the network model size only increases by 0.059MB, andthe reasoning speed is remained unchanged, it is easier to deploy in embedded devices for real production activities with the same accuracy comparable to that of the standard structured networks.

Key words: Myrica rubra, YOLOX-NANO algorithm, ECA, Focal Loss, EIoU

摘要： 为实现杨梅采摘智能化，开发杨梅成熟度检测设备，提出一种基于改进YOLOX-NANO算法的杨梅果实成熟度检测方法。通过在特征加强提取网络层中引入通道注意力模块，提高网络对通道特征的提取能力；引入焦点损失函数代替标准交叉熵损失函数，解决单阶段网络正负样本不均衡问题，避免梯度方向指向非最优解；使用高效交并比损失函数，提高网络模型对目标识别的准确率。试验结果表明，在自建数据集上与原YOLOX-NANO相比，改进YOLOX-NANO算法对于三种不同成熟度杨梅果实的识别精度均有提升，平均精度达到92.67%，而网络模型大小只增加0.059MB，推理速度不变，在精度达到与标准结构网络相当的前提下，更易于部署到嵌入式设备中。

关键词: 杨梅, YOLOX-NANO算法, 通道注意力机制, 焦点损失函数, 高效交并比