基于轻量卷积网络的田间自然环境杂草识别方法

Weed identification method based on deep transfer learning in field natural environment

针对田间自然环境下杂草识别精度低和检测速度慢的问题,本文依据自然环境杂草图像数据的特性,在Xception卷积网络的基础上构建了一种基于轻量卷积网络的杂草识别模型。首先改进Xception模型,采用ELU作为模型的激活函数,并使用全局最大池化层对最后一层卷积进行下采样。然后,对原始数据进行背景分割和数据增强处理,在迁移后的模型上继续微调,训练得到最佳的杂草识别模型。在相同的试验条件下,与VGG16、VGG19、ResNet50和Inception-V3四种标准的深度卷积网络模型进行比较,结果显示,本文模型的整体性能最好,对自然条件下8类杂草及苗期玉米的平均测试识别准确率高达98.63%,改进模型的规模为83.5 MB,单张杂草图像检测平均耗时仅为63.8 ms。本文研究结果可为田间自然环境下精准喷药的实施提供理论基础和技术支持。

At present, weed identification mostly uses weeds images taken in laboratory, and it is difficult to quickly and accurately identify weeds in a complex field environment. In order to improve low weed identification accuracy and slow detection speed in natural field environments, according to the characteristics of the weed image data, this study proposed a weed recognition model based on Xception lightweight convolutional network by improving the activation function and pooling layer. Firstly, the Xception was improved using Exponential Linear Unit(ELU) as the activation function of the model, and the global max pooling layer was used to downsample the convolution of the last layer. Then, the weed image was processed by background segmentation and data enhancement, and the best model was obtained by fine tuning on the pre-trained model. Compared with the four standard deep convolutional network models of VGG16, VGG19, ResNet50 and Inception-V3 under the same experimental conditions, the experimental results show that the overall performance of the model in this paper was optimal with the test accuracy rate as high as 98.63%. The scale of the model was 83.5 MB, and the average time for detecting a single weed image was only 63.8 ms. This study can provide theoretical basis and technical support for the implementation of precision spraying in natural field environment.