基于多层信息融合和显著性特征增强的农作物病害识别

Crop Disease Recognition Based on Multi-layer Information Fusion and Saliency Feature Enhancement

现有基于深度学习的农作物病害识别方法对网络浅层、中层、深层特征中包含的判别信息挖掘不够,且提取的农作物病害图像显著性特征大多不足,为了更加有效地提取农作物病害图像中的判别特征,提高农作物病害识别精度,提出一种基于多层信息融合和显著性特征增强的农作物病害识别网络(Crop disease recognition network based on multi-layer information fusion and saliency feature enhancement,MISF-Net)。MISF-Net主要由ConvNext主干网络、多层信息融合模块、显著性特征增强模块组成。其中,ConvNext主干网络主要用于提取农作物病害图像的特征;多层信息融合模块主要用于提取和融合主干网络浅层、中层、深层特征中的判别信息;显著性特征增强模块主要用于增强农作物病害图像中的显著性判别特征。在农作物病害数据集AI challenger 2018及自制数据集RCP-Crops上的实验结果表明,MISF-Net的农作物病害识别准确率分别达到87.84%、95.41%,F1值分别达到87.72%、95.31%。

Crop disease recognition is a prerequisite for rational pesticide application and a powerful guarantee for promoting healthy and stable agricultural development.Existing deep learning-based crop disease recognition methods mainly use classical networks such as VGG and ResNet or networks that use attention mechanisms for disease recognition.Although these deep learning-based crop disease recognition methods have achieved better disease recognition results than traditional methods,they do not sufficiently mine the discriminative information contained in the shallow,middle and deep features of networks,and most of the extracted saliency features of crop disease images are insufficient.To extract discriminative features in crop disease images more effectively and improve crop disease recognition accuracy,a crop disease recognition network based on multi-layer information fusion and saliency feature enhancement(MISF-Net)was proposed.Specifically,MISF-Net mainly consisted of a ConvNext backbone network,a multi-layer information fusion module(MIFM),and a saliency feature enhancement module(SFEM).The ConvNext backbone network was mainly used to extract features of crop disease images.The multi-layer information fusion module was mainly used to extract and fuse the discriminative information from the shallow,medium and deep layers of the backbone network.The saliency feature enhancement module was mainly used to enhance the saliency discriminative features in crop disease images.The experimental results on the crop disease dataset AI challenger 2018 and the homemade dataset RCP-Crops showed that the crop disease recognition accuracies of MISF-Net reached 87.84%and 95.41%,and the F1 values reached 87.72%and 95.31%,respectively.