【摘要】 针对枸杞产业现代化发展中枸杞高效低损智能化采摘的需求，项目组提出一种基于深度学习方法的成熟枸杞在线检测方法。首先，通过互联网和实地拍摄完成原始图像的收集，对图像预处理后通过噪声添加、角度旋转等操作完成数据集扩展；其次，对YOLOv5模型进行训练、参数优化和检测试验，并与一阶段和两阶段算法进行比较。结果表明：在不同背景、光照、天气、相机位置等情况下，对成熟挂枝枸杞的平均识别准确率为88%，召回率高达96%；单幅图片的平均测试时间为7.9 ms，红果和绿果区分明确，误检率极低；在轻微遮挡的情况下，漏检率低。同时，试验得出YOLOv5的检测精度和实时性在所有对比算法中性能最佳，满足枸杞采摘的市场需求，为高效低损智能化的枸杞采摘提供了可靠的算法支持。更多还原

【Abstract】 Aiming at the demand of high efficiency and low-loss intelligent picking of wolfberry for the modernization of wolfberry industry, this paper proposes a method of online detection of red wolfberry based on deep learning. First, the original images acquisition was completed by Internet and field photography, and made the data set by noise addition, angle rotation and other ways after images pre-processing; Then the YOLOv5 model was trained, parameter optimized, tested, and compared with one-stage and two-stage algorithms. The results showed that the average recognition accuracy of red wolfberry was up to 88%, the recall rate was 96%, the average detecting time of a single image was 7.9 ms, the distinction between red and green fruits was clear,basically no false detection, and the leakage rate was low in case of slight occlusion. The results also showed that the detection accuracy and real-time performance of YOLOv5 was the best among all the compared algorithms, which met the market demand of wolfberry picking and provides reliable algorithm support for efficient and low-loss intelligent wolfberry picking.更多还原