Cassava Disease Classification Based On VGG16 Network

Cassava is a significant crop in China,so how to improve its yield is of particular concern to growers.According to research,the disease is the most important factor affecting cassava yield.Early diseases of cassava are usually reflected by the state of its leaves,and therefore it is of practical interest to leverage artificial intelligence technologies to accurately classify diseases of cassava leaves.In recent years,how to use convolutional neural networks to accurately identify the type of diseases of cassava leaves has become a hot research topic among scholars.This project is to classify the diseases of cassava leaves based on the VGG16 network,and the main research contents of the project are as follows:(1)To address the problem of the insufficient data set and training samples for the classification of cassava leaf diseases,this paper uses the strategy of transfer learning.Conducting the recognition task through DenseNet,MobileNet,ResNet,and VGG16 models that have been pre-trained for the task.As a result,the best-performing VGG16 model achieved an accuracy of 88.5%on the test set.Compared with VGG16 without transfer learning,an improvement of 10%was obtained in accuracy.(2)Based on transfer learning,the paper fuses VGG16 with three pre-trained models,DenseNet,MobileNet,and ResNet,respectively,to obtain three different joint models based on VGG16.In the stage of feature fusion,the paper proposes a feature fusion strategy based on the self-attention mechanism,which enables the joint model to focus on the key features selectively and dynamically.VGG16-DenseNet performs the best among the three joint models,achieving an accuracy of 90.2%,an increase of 1.7%compared to the pre-trained VGG16 model alone.(3)In order to further improve the performance of the joint model on the cassava leaf disease classification task,the paper optimizes the joint model in terms of focal loss,hierarchical learning rate,and learning rate warmup,respectively,and designs experiments to demonstrate that the accuracy of the joint model can be further improved in these three optimization algorithms.As a result,the joint model VGG16-DenseNet designed in the paper achieves 92.3%and 87.7%accuracy respectively on the classification test set and extended data set of cassava leaf images.