基于混合注意力机制的番茄叶片病害识别方法

A Tomato Leaf Disease Recognition Method Based on Hybrid Attention Mechanism

番茄叶片病害对番茄的产量和质量构成了严重威胁,为了实现对番茄病害的早期识别和预防,因此需要有效的番茄病害识别方法。香农熵可以作为衡量图像信息量大小的度量,但在深度卷积神经网络中通道香农熵的计算量巨大导致训练时间增长,由于图像香农熵和图像方差之间存在正相关性,因此可以通过计算图像方差来衡量图像的信息量。在深度残差卷积神经网络基础上提出了基于Squeeze-and-Excitation(SE)和通道方差的混合注意力机制深度卷积神经网络方法。先计算通道的方差数据以此衡量通道信息量大小,方差大的通道信息量大反之亦然,根据通道方差值对通道进行加权处理,再对通道进行SE操作。实验表明:提出的方法能有效地对番茄病害类型进行分类和识别,与传统方法相比准确率提高了4.65%。

The disease of tomato leaf becomes a serious threat to the yield and quality of tomatoes.In order to achieve early detection and prevention of tomato diseases,it is necessary to have an effective method for identifying tomato diseases.Shannon entropy can be used as a measure ment of the amount of information in an image,but a huge calculation of channel Shannon entropy in deep convolutional neural networks can increase the training time of network.Since there is a positive correlation between image Shannon entropy and image variance,image variance can be calculated to measure the amount of information in an image.This paper proposes a deep convolutional neural network method based on the Squeeze-and-Excitation(SE)mechanism and channel variance.Firstly,the variance data of the channels are calculated to measure the channel information,which a big variance of channel has more information,and vice versa.According to the variance values of the channels,weighted processing of the channels is carried out,followed by SE operations on the channels.The experiments show that the method proposed in this paper can effectively classify and identify types of tomato diseases,with an accuracy improvement of 4.65%compared to traditional methods.