利用轻量型卷积神经网络模型识别苹果叶部病害的探索

Exploration of identifying apple leaf diseases using lightweight convolutional neural network model

网络深度和宽度的增加虽然能增加识别的准确率,但往往参数量和计算量较大,并不适合移动设备的应用。为解决这一问题,构建了2种轻量型CNN模型,通过改进SqueezeNet网络的Fire模块,加入空间注意力机制以及在网络深层加入稠密连接模块,提高网络的特征提取与特征复用能力。通过在构建的苹果病害叶片数据集上训练,改进后的模型识别准确率达到89.60%和94.37%,相较于原网络提高了2.98个和7.75个百分点,而网络的参数量仅有0.9 M和2.5 M。结果表明,改进后的网络在保证模型轻量的同时也获得了较高的识别准确率。

Although the increase of network depth and width can enhance the recognition accuracy,it is often not suitable for mobile device applications due to the large number of parameters and calculations required.To address this problem,we developed two lightweight CNN models.These models improve the feature extraction ability of the network by enhancing the Fire module of the SqueezeNet network,incorporating a spatial attention mechanism,and introducing a dense connection module in the deep layer within the network.Through training on an apple disease leaf dataset,the recognition accuracy of the improved model reaches 89.60%and 94.37%,exceeding the original network's accuracy by2.98%and 7.75%,respectively.Remarkably,the number of parameters in these networks remains low,at only 0.9M and2.5M.The experimental results showed that the improved network not only maintains a lightweight model but also achieves higher recognition accuracy.