面向番茄病害识别的改进型SqueezeNet轻量级模型

An Improved SqueezeNet Lightweight Model for Tomato Disease Recognition

针对番茄病害识别中深度神经网络参数过多、识别精度较低的问题,从网络轻量化和提取特征精准化的角度出发,对SqueezeNet结构进行改进。为精简fire模块,对其中Expand层的卷积核大小、网络层数以及通道数进行调整。同时,将模型与ECA模块结合,利用局部跨通道交互的方式获得各通道的注意力值,强化网络对关键特征的学习能力。实验结果表明,与LeNet、MobileNet和SqueezeNet模型相比,改进型SqueezeNet的模型大小和识别准确率均具有明显优势,为嵌入式设备在实际生产中的番茄病害识别提供一种技术方法。

Aiming at the problems of too many parameters and low recognition accuracy in deep neural network for tomato disease recognition,the SqueezeNet structure was improved from the perspective of simplified network and precise features recognition.In order to simplify the fire module,the size of the convolution kernel,the number of network layers and channels in the Expand layer were adjusted.At the same time,the model was combined with the ECA module to obtain the attention value of each channel by local cross-channel interaction,so as to strengthen the network′s learning ability for key features.The experimental results showed that the improved SqueezeNet had obvious advantages in model size and recognition accuracy compared with the LeNet,MobileNet and SqueezeNet models,which provided a technical method for tomato disease recognition in application of embedded devices.