结合图像处理和深度学习的麦穗识别方法

Wheat Head Detection Method Based on Image Processing and Deep Learning

摘要: 麦穗识别在智慧农业中有较高的应用价值，如粮食产量估计、种子筛选、和种子基因性能等。许多研究人员主要关注麦穗计数。传统的大田麦穗计数具有简单、方便等特点，但效率低下、人的主观因素较多。本研究采用Global Wheat Head detection (GWHD) dataset的小麦图像数据。为了准确对数据集进行分类识别，借助图像的基本操作实现数据增强，而后在结合深度学习在图像识别中的优势，以ResNet50为基础网络框架，设计一套基于Faster R-CNN的麦穗图像识别算法。实验结果表明，相比其他卷积神经网络，ResNet50网络模型具泛化性好，鲁棒性强、精度高等优点。

Abstract: The automatic detection of wheat head has high application value in yield prediction, seed screening, density estimation and gene character expression. At present, the main research focuses on wheat head counting. The traditional counting method of wheat head is simple and convenient, but it has the problems of low efficiency and high subjective influence. Wheat image data comes from Global Wheat Head detection (GWHD) dataset in this study. To accurately classify and recognize the data set, data enhancement is realized with the help of the basic operation of the image. Then, combined with the advantages of deep learning in image recognition, a set of wheat head image recognition algorithm based on Fast R-CNN is designed based on ResNet50 network framework. The experimental results show that compared with other convolutional neural networks, ResNet50 network model has the advantages of good generalization, strong robustness, and high accuracy.