一种基于VovNet的轻量级农作物虫害分类模型

A Lightweight Crop Pest Classification Model based on VovNet

由于传统CNN模型参数量大,对训练样本及算力要求高,利用深度学习进行农作物虫害识别容易造成农作物识别受到硬件条件限制。本文在VovNet基础上设计一个轻量级LVovNet模型,将VovNet的普通卷积替换为深度可分离卷积,减少了模型参数,提高了GPU利用率,在模型最后加入归一化通道注意力机制,加强网络特征提取能力并控制了网络参数量。用盲蝽、蝗虫、红蜘蛛等12类常见农作物虫害类别共5 785张RGB图片作为测试数据进行验证,该模型识别精度达97.34%,与VGG、ResNet、DenseNet、VovNet等相比,具有参数少、复杂度低、网络延迟低、识别精度高等特征。

Due to the large number of parameters of the traditional CNN model and the high requirements for training samples and computing power,the use of deep learning for crop pest identification is likely to cause crop identification to be limited by hardware conditions.In this paper,a lightweight LVovNet model is designed based on VovNet.The ordinary convolution of VovNet is replaced by deep separable convolution,which reduces the model parameters and improves the GPU utilization.At the end of the model,the normalized channel attention mechanism is added to strengthen the network feature extraction ability and control the number of network parameters.A total of 5785 RGB images of 12 common crop pests categories such as mirids,locusts and red spiders were used as test data.The recognition accuracy of the model was 97.34%.Compared with VGG,ResNet,DenseNet and VovNet,it has the characteristics of less parameters,low complexity,low network delay and high recognition accuracy.