基于超轻量化卷积神经网络的番茄病虫害诊断

Diagnosis of tomato pests and diseases based on super lightweight convolutional neural network

针对番茄病虫害诊断中存在的传统卷积神经网络结构复杂、难以直接应用于便携终端,以及现有轻量化卷积神经网络特征提取能力弱、识别准确率低、难以满足实际需要等问题,本研究拟在原有轻量化卷积神经网络的基础上,定义超轻量化卷积神经网络,设计一种基于SqueezeNet网络改进的超轻量化卷积神经网络,将其用于番茄病虫害诊断任务中。首先,改进SqueezeNet网络中的Fire模块,生成2种适用于不同特征维度的Fire模块,并引入ECA(高效通道注意力)模块以提高模型的特征提取能力;其次,结合扩展型指数线性单元函数(SELU)和Mish函数,替代修正线性单元函数(ReLU)作为激活函数;再次,采用软池化(Softpool)替代原始的最大池化;最后,利用中心损失函数(Center loss)改进指数归一化损失函数(Softmax loss),提高对近似病虫害的识别准确率。本研究选择了8种害虫和9种病害,对害虫、病害、病虫害3类数据集进行数据增强,并探讨了数据的小样本性、不平衡性对模型性能的影响。结果表明,本研究提出的模型具有超轻量化的特点,对害虫、病害、病虫害的识别准确率最高分别可达98.83%、98.14%和97.71%,能够很好地满足番茄病虫害诊断需求。

In the diagnosis of tomato diseases and pests,traditional convolutional neural network structures are complex and hard to be directly applied to portable terminals.Besides,existing lightweight convolutional neural networks exhibit weak feature extraction capabilities,low recognition accuracy,and are inadequate for practical applications.Aiming at the above problems,we intended to define a super lightweight convolutional neural network based on existing lightweight convolutional neural network,and to design an ultra-lightweight convolutional neural network by improving the SqueezeNet network for tomato disease and pest diagnosis tasks.Firstly,we enhanced the Fire module in the SqueezeNet network,generated two Fire modules suitable for different feature dimensions.We introduced efficient channel attention(ECA)module to improve feature extraction capabilities of the model.Secondly,we incorporated scaled exponential linear unit(SELU)and Mish to replace rectified linear unit(ReLU)as activation function.Next,we employed Softpool instead of the original max pooling.Finally,we enhanced the exponential normalized loss(Softmax loss)by using Center loss function to improve the recognition accuracy of approximate diseases and pests.In this experiment,we selected eight types of pests and nine types of diseases to perform data augmentation on three datasets(pests,diseases,diseases and pests),and investigated the impact of small sample and data imbalance on model performance.Experimental results demonstrated that the network proposed in this study had super lightweight characteristics.The recognition accuracies for pests,diseases,and diseases and pests could reach up to 98.83%,98.14%and 97.71%,respectively,which met the requirements for diagnosis effectively.