一种基于迁移学习的玉米发育期识别方法研究

A Recognition Method of Corn Development Stage Based on Transfer Learning

摘要: 本文利用深度学习模型迁移方法，选择基础模型瓶颈层的输出作为提取的特征，后面训练的层作为分类器，实现玉米图片发育期的自动分类识别。在深度学习框架TensorFlow下，分别搭建基于inception V3和vgg16的迁移网络，通过调整模型输入大小，搭建了基于inception V3原始尺寸、基于inception V3大尺寸和基于vgg16原始尺寸、基于vgg16大尺寸四种不同的模型。在玉米数据集上进行模型训练、测试，对比四种模型的训练精度和检测精度。结果表明针对玉米发育期识别问题，同一种基础模型搭建的迁移网络，不同尺寸模型，训练精度差别不大，测试精度大尺度模型明显优于原始尺寸模型。基于inception V3大尺寸的迁移网络，模型深，参数多，并且模型占用空间小，测试准确率高，更适合本文玉米图片的发育期分类识别。

Abstract: Using the deep learning transfer method, the study selects the output of the bottleneck layer of the basic model as the extracted features, and uses the later training layer as the classifier to realize the automatic classification and recognition of the images of corn development stages. Under the deep learning framework of TensorFlow, the transfer networks based on inception V3 and vgg16 are built respectively. By adjusting the model input size, we build four different models, which are based on the original size of inception V3, the large size of inception V3, the original size of vgg16 and the large size of vgg16. The model was trained and tested on corn data set to compare the training accuracy and detection accuracy of the four models. On the recognition of corn development stages, the transfer network constructed by the same basic model and different size models had little dif-ference on training accuracy, and the large-scale model was significantly better than the original size model on testing accuracy. Due to its features of deeper model, more parameters and taking up less space, the large-scale transfer network based on inception V3 with higher test accuracy is more suitable for the classification and recognition of corn pictures in this paper.