基于深度学习的林下落果识别方法与试验

监测与识别林下落果的数量和分布信息,是实现落果自动收获和果园智能化管理的重要基础。针对目前落果识别智能化程度较低等问题,提出一种基于深度学习的林下落果识别方法。首先,以不同类型、品种落果图像为基础,通过数据预处理、增强等方法建立林下落果图像数据集。其次,利用YOLO-v3深度卷积神经网络优势特性,建立落果智能识别方法。最后,以柑橘、梨、苹果三种典型落果,对基于深度学习的林下落果识别方法进行测试与验证,分析相关试验结果。试验结果表明:所提出的基于YOLO-v3落果识别方法,在不同条件均能准确识别落果,三种典型落果识别精度大于89%;相对于SSD,RCNN和CenterNet三种网络模型,YOLO-v3的准确率分别提高7%,2%和3.5%;在腐烂落果识别层面,YOLO-v3、SSD、RCNN和CenterNet的识别准确率分别为86%,59%,64%和43%;YOLO-v3的识别准确率高于其他深度学习模型。所提出的方法可以精确的识别林下落果,为后期的落果精准管理提供必要的技术支撑。

Identification method and experiment of fallen fruit based on deep learning

In order to monitor the quantity and distribution and provide the necessary support for automatic harvesting of the fallen fruit and intelligent orchard management, this paper proposes a deep learning based fruit fall identification method to solve the problem of low automatic monitoring degree of fruit fall at present. Firstly, the fallen fruit image dataset was established with consideration of different images by using data collection and pre processing. Secondly, YOLO v3 (You Only Look Once) deep convolutional neural networks were used to identify the fallen fruits. Finally, an experiment was conducted toevaluate the performance of the proposed method. The results showed that the proposed algorithm could accurately identify fallen fruit under different types and different occlusion conditions. YOLO v3 got the most value (more than 89%) of the accuracy of indentation. Compared with SSD, RCNN, and CenterNet, the accuracy of YOLO v3 increasedaccuracy by 7%, 2%, and 3.5%, respectively. The recognition accuracy of rotting fallen fruits by using YOLO v3, SSD, RCNN, and CenterNetreached 86%, 59%, 64%, and 43%, respectively. The recognition accuracy of YOLO v3 was higher than that of other deep learning models. The proposed method can accurately identify the number of the fallen fruit and provide necessary technical support for accurate management of fallen fruit.