基于轻量化卷积神经网络的油茶病害识别OA北大核心CSTPCD

[Objective]A multi-scale Camellia oleifera Leaf Disease Recognition Net(COLDR-Net)based on improved lightweight convolutional neural network ShuffleNet V2 was proposed to address the issue of low accuracy in image recognition of Camellia oleifera leaf diseases in natural environments.[Method]The model enhanced the disease spot feature information in the image with the embedded Efficient Channel Attention(ECA).Additionally,a Multi-scale Feature Extraction(MEF)unit was designed to enhance the recognition ability of fine spots.Focal Loss was used to replace the Cross-Entropy Loss function,which alleviated the unbalanced distribution of sample categories leading to large differences in the effectiveness of the model for different categories of disease identification.The Mish activation function was used to avoid the problem of gradient disappearance when the input was negative and to improve the expression ability of the model.The network structure was optimized by trimming the number of network layers and optimizing the number of output channels to reduce the number of computations and model parameters to achieve the lightweight of the model.[Result]The results showed that the accuracy and F1-score of COLDR-Net on the disease dataset of Camellia oleifera leaf were 97.19%and 97.08%,respectively.Its accuracy was higher when compared with that of AlexNet(93.04%),VGG16(94.18%),ResNet18(94.5%),ResNet50(95.45%)and MobileNetV3-Large(93.41%).The accuracy was 4.07%better than the model before improvement.The number of model parameters were 2.61M,the FLOPs were 0.24G,and the average inference time for a single image on the mobile devices were 67ms.The model was deployed on the mobile Android platform to develop a leaf disease identification system for Camellia oleifera.[Conclusion]The proposed COLDR-Net model effectively meets the demand for real-time identification of Camellia oleifera leaf diseases,which can provide valuable reference for disease control,diagnosis,and resource-constrained mobile terminals.