基于改进ResNet34的水稻叶片病害识别模型

打开文本图片集

关键词：水稻叶片病害；深度可分离卷积层；非局部注意力机制；深度学习；卷积神经网络中图分类号：S511；S435.11；TP391.4 文献标识码：A 文章编号：2095-5553（2025）12-0086-08

Abstract：Riceleafdiseasescanseriouslyafectriceyield，soitisvery importanttoidentifyriceleafdiseasestimelyand accurately.Convolutionalneuralnetworksperformwellinthefieldofimagerecognition，butforriceleaf disease recognition task，theexisting CNNmodelhas theproblemofinsufficientfeatureextractionability.Inthispaper，arice leaf disease recognition model based on improved ResNet34 is proposed.Thefeatureextraction capabilityof the model is enhancedbyintroducingnon-localatentionmechanismanddeepseparableconvolutionlayer.Thenon-localatention mechanism effectivelyfocuses onthe Keyareas relatedtothediseasein theimage，which improves thesenstivityof the model tothe disease features.The depth separable convolution layer significantly reduces the numberof model parameters whilemaintainingtheperformanceoffeatureextraction.Theefectivenessoftheproposedalgorithmisverifiedby comparing it with ResNet34，ResNet101，VGGl6，VGG19，MobileNetV2 and Swin Transformeron rice leaf disease data set. Compared to the ResNet34 model，the algorithm improved accuracy by 5.6% ，recall rateby 4.8% ，and F1 score by 5.2% . Inaddition，the robustness of thealgorithm is evaluated，and the results show that thealgorithm has good robustness to image rotation，scaling and noise.

Keywords：riceleaf disease；deep separable convolution layer;non-local atention mechanism；deep learning; convolutional neural networks

0 引言

水稻作为全球主要的粮食作物之一，其健康生长对于保障世界粮食安全至关重要[1]。（剩余12506字）