Abstract: In order to achieve fast and accurate detection of macadamia nuts in natural environment, a macadamia detection algorithm based on SCGYOLOv5n during the harvesting period is proposed, aiming at the problems of the similar color of macadamia nut peel and branch leaves during harvesting period, small size and difficult identification of mixed diseased fruit. The method uses data augmentation to improve model robustness, introduces SimAM attention mechanism in the backbone network of YOLOv5n to enhance the extraction of effective features, introduces CARAFE up sampling in the FPN structure to strengthen target perception, uses GSConv lightweight convolution to replace some convolutional layers to reduce the number of model parameters and achieve efficient feature fusion to improve detection speed and detection accuracy. The results show that the improved SCGYOLOv5n macadamia detection algorithm has an average accuracy AP of 94.8% and 97.9% for the detection of green macadamia nuts and diseased macadamia nuts during the harvest period, respectively, and the average time of a single image is 5.33ms, which is 2.1% and 1.3% higher than the YOLOv5n model, and the detection speed is improved by 15.8%. The algorithm can efficiently detect macadamia nuts and provide technical reference for subsequent automated harvesting.

Key words: deep learning,  , macadamia nut detection, data augmentation, attention mechanism, convolutional neural network design

摘要： 为实现自然环境下澳洲坚果的快速准确检测，针对收获期澳洲坚果果皮与枝叶颜色相似、体积小、病害果混杂难识别的问题，提出一种基于SCGYOLOv5n的收获期澳洲坚果检测算法。该方法运用数据增强，提高模型鲁棒性；在YOLOv5n的骨干网络引入SimAM注意力机制，增强有效特征的提取能力；在FPN结构中引入CARAFE上采样，强化目标感知能力；使用GSConv轻量级卷积替换部分卷积层，减轻模型的参数量并实现高效特征融合，提高检测速度和检测精度。结果表明，改进后的SCGYOLOv5n澳洲坚果检测算法对收获期的青皮澳洲坚果和病害澳洲坚果的检测平均精度AP分别为94.8%、97.9%，单张图像平均时间为5.33 ms，比YOLOv5n模型高出2.1%、1.3%，检测速度提升15.8%。该算法可以高效检测澳洲坚果，为后续自动化采摘提供技术参考。

关键词: 深度学习, 澳洲坚果检测, 数据增强, 注意力机制, 卷积神经网络设计