基于三通道注意力网络的番茄叶部病害识别

Tomato Leaf Disease Recognition Based on Three-channel Attention Network

对番茄病害进行识别,近年来一直是植物病害预防的研究热点。由于受到复杂背景干扰,番茄叶部病害识别准确率不高,针对这一问题,提出一种基于三通道注意力机制网络的番茄叶部病害识别方法。该网络基于ResNeXt50残差网络,将注意力模块嵌入至残差网络的ResNeXt模块中可以并行提取目标的通道特征和空间特征,获取有效的语义信息。训练阶段通过设计双损失函数和数据增强进一步提升分类准确度,并通过迁移学习网络预训练参数的方式提高网络训练效率。实验结果表明,使用双损失函数和数据增强后,基于三通道注意力网络的番茄病害识别算法在测试集上的平均识别准确率达98.4%,相比于传统机器学习方法和其他神经网络方法的准确率更高,检测速度满足实时性,Kappa系数为0.96,满足叶部病害识别的高精度要求。该方法能够有效地对10种番茄叶部病害进行识别,为植物病害识别提供了一种新的思路。

The identification of tomato diseases has been a research hotspot in plant disease prevention in recent years.Due to the interference of complex background,the accuracy of tomato leaf disease recognition is not high.To solve this problem,a tomato leaf disease recognition method based on a three-channel attention mechanism was proposed.Based on the ResNeXt50 residual network,the attention module was embedded in the ResNeXt module of the residual network to extract the channel features and spatial features of the target in parallel,and obtain effective semantic information.In the training stage,double loss function and data enhancement were designed to further improve the classification accuracy,and the network training efficiency was improved by means of transfer learning network pre-training parameters.The experimental results show that the average recognition accuracy of the tomato disease recognition algorithm based on three-channel attention network reaches 98.4%on the test set after using dual loss function and data augmentation,which is more accurate than traditional machine learning methods and other neural network methods,and the detection speed satisfies real-time with a Kappa coefficient of 0.96,which meets the high accuracy requirements for leaf disease recognition.The method can effectively identify 10 kinds of tomato leaf disease,which provides a new idea for plant disease identification.