基于改进卷积神经网络的辣椒病虫害检测

Pepper pest detection based on neural network

针对使用卷积神经网络对辣椒病虫害进行检测有参数多、计算量大和推理时间过长等问题,提出一种基于MobileNet-V2改进的轻量化神经网络,将MobileNet-V2的BN层中的激活函数全部替换为Leaky ReLU,保留特征图中更多的有效正负信息,以提高性能和减少计算复杂度,增强模型的鲁棒性。在公开的辣椒病虫害数据集上使用VGG16、ResNet34和MobileNet-V2等模型对比后,改进的MobileNet-V2表现出更高的准确性和更少的参数量。相对于原来的MobileNet-V2准确率提升4%,相对VGG16、ResNet34两种模型参数分别下降97%和87%。能够移动端设备实现实时病虫害检测,提供高效便捷解决方案。

There are many problems in using convolutional neural network tOdetect pepper diseases and insect pests,such as large number of parameters,large amount of calculation and toOlong inference time.This paper proposes an improved lightweight neural network based on MobileNet-V2,replacing all activation functions in the BN layer of MobileNet-V2 with Leaky ReLU,retaining more effective positive and negative information in the feature map toimprove performance and reduce Computational complexity and enhanced model robustness.After comparing models such as VGG16,ResNet34 and MobileNet-V2 on the public pepper diseases and insect pests data set,the improved MobileNet-V2 showed higher accuracy and fewer parameters.Compared with the original MobileNet-V2,the accuracy increased by 4%,and compared with VGG16 and ResNet34,the parameters of the twOmodels dropped by 97%and 87%respectively.It can realize real-time pest and disease detection on mobile devices and provide efficient and convenient solutions.