基于全局自注意力的小麦图像识别

Wheat Image Recognition Based on Global Self-attention

在实际应用场景下,通过图像识别的方式来识别小麦的病虫害具有极大的挑战性。与以往纯粹基于卷积神经网络(Convolutional Neural Network,CNN)的方法相比,将小麦图像转换成一系列视觉语言,并从全局视角进行小麦识别的方法是更可行和实用的。运用Convolutional Visual Transformers(CVT)来解决小麦识别分为2个环节。首先,利用2分支CNN生成的2种特征图来实现注意选择性融合(Attentional Selective Fusion,ASF)。ASF通过融合多个特征和全局-局部注意力来获取有区别的信息,并投射成一系列的视觉语言。其次,受Transformers在自然语言处理方面的成功启发,用全局自注意力来建模这些视觉语言之间的关系。将CVT与经典分类网络LeNet-5、ResNet-18、VGG-16、EfficientNet对比,识别率有所提升,同时该方法具有良好的泛化能力。

In the actual application scenario,it is very challenging to identify wheat diseases and pests by image recognition.Compared with the previous methods based solely on convolutional neural network(CNN),the method of converting wheat images into a series of visual languages and recognizing wheat from a global perspective is more feasible and practical.The use of convolutional visual Transformers(CVT)to solve wheat recognition is divided into two links.First,two feature maps generated by two-branch CNN are used to realize attentional selective fusion(ASF).ASF obtains different information by fusing multiple features and global-local attention,and projects it into a series of visual languages.Secondly,inspired by the success of Transformers in natural language processing,global self-attention is used to model the relationship between these visual languages.Compared with classical classification networks LeNet-5,ResNet-18,VGG-16 and EfficientNet,CVT improves the recognition rate,and this method has good generalization ability.