基于小型卷积神经网络的南疆棉花图像分类

针对南疆棉田图像分类特定应用场景,该文提出一种小型卷积神经网络模型,分别对幼苗期棉苗、缺苗穴和地膜及吐絮期叶片、成铃和吐絮铃图像进行自动化分类。该卷积神经网络由12层组成,包括交替堆叠的4个卷积层和4个最大池化层,以及1个展平层、1个Dropout层和2个密集连接层。采用智能手机拍照方法,获取幼苗期棉苗、缺苗穴和地膜图像13 920张,吐絮期叶片、成铃和吐絮铃图像21 427张。在普通笔记本电脑上部署TensorFlow、Keras深度学习框架和卷积神经网络模型,使用数据增强和添加Dropout层来消除过拟合。研究结果表明,小型卷积神经网络在幼苗期和吐絮期图像测试集的分类精度分别达到了0.999 3和0.975 7,模型具有很好的泛化能力,模型的训练时间约2 h。研究结果将为利用数字图像智能提取棉花缺苗信息及棉铃吐絮信息提供一定的参考。

Aiming at the specific application scene of cotton field image classification in southern Xinjiang,a small convolution neural network model is proposed to automatically classify the images of cotton seedlings,lack of seedling holes and plastic film in seedling stage,as well as leaves,bolls and bolls in boll opening stage.The ConvNet consists of 12 layers,including 4 convolution layers and 4 maximum pooling layers stacked alternately,as well as 1 flattening layer,1 Dropout layer and 2 dense connection layers.Total number of 13920 images of cotton seedlings,seedling deficiency points and films at seedling stage,and 21427 images of leaves,mature bolls and opening bolls at boll opening stage were obtained by taking pictures with smart phones.TensorFlow and Keras deep learning framework and convolution neural network models on ordinary laptops were constructed by using data augmentation and adding Dropout layers to get rid of overfitting.The results showed that the classification accuracy of test set at seedling stage and boll opening stage reached 0.9993 and 0.9757,respectively.The model had good generalization ability,and the training time of the model was about 2 hours.The research results will provide some reference for the intelligent extraction of cotton seedling deficiency information and boll opening information by using digital images.