基于YOLOv5s的神经网络麦穗识别算法研究

Research on Wheat Ears Recognition Algorithm Based on YOLOv5s Neural Network

摘要: 神经网络算法在人工智能领域有着广泛的应用，如本文中智慧农业中的麦穗识别。本研究采用YOLOv5s 算法研究麦穗的目标检测，将目标检测算法和神经网络算法相结合。研究的数据集来自Global Wheat Head detection (GWHD) dataset2020和AI crowd中GWHD dataset 2021，利用LabelImg软件进行标注，将标注后的数据先采用PIL中的旋转、翻转、亮度提升等对数据进行增强预处理，其次，选用YOLO系列算法对预处理后的数据集进行目标检测。增大麦穗训练集的规模，提高模型的泛化能力。通过YOLO算法进行麦穗检测，确保对数据集进行目标检测的准确性。利用该模型对麦穗图像进行检测。其中YOLOv5算法在输入端增加了Mosaic数据增强、自适应anchor以及图像缩放板块。并在neck上使用FPN和PAN结合板块，其精确率、召回率和平均精度在测试集上的精度分别为78.1%、85%、52.1%，比YOLOv3的检验结果高，该方法可以高效地检测并标注麦穗。

Abstract: Neural network algorithm has a wide range of applications in the field of artificial intelligence, such as wheat ear recognition in Intelligent Agriculture in this paper. In this study, YOLOv5s algorithm is used to study the target detection of wheat ears, which combines the target detection algorithm with neural network algorithm. The data sets studied are from global wheat head detection (GWHD) dataset 2020 and GWHD dataset 2021 in AI crowd. LabelImg software is used to label the marked data. Firstly, the rotation, turnover and brightness improvement in PIL are used to enhance the preprocessing of the data. Secondly, YOLO series algorithms are selected to detect the target of the preprocessed data set, increase the scale of wheat ear training set and improve the generalization ability of the model. The ear of wheat is detected by YOLO algorithm, and the target of the data set is detected accurately. The model is used to detect the wheat ear image. Among them, YOLOv5 algorithm adds mosaic data enhancement, adaptive anchor and image scaling plates at the input. Using FPN and pan combined plates on neck, the accuracy, recall and average accuracy on the test set are 78.1%, 85% and 52.1% respectively, which is higher than the test result of YOLOv3. This method can detect and label wheat ears efficiently.