Application Of Deep Learning In Cassava Leaf Disease Classification

This thesis studied the application of deep learning methods in cassava leaf disease classification.Cassava is the main food crop in Sub-Saharan Africa,but the rapid spread of its diseases has seriously endangered the yield,which is essential to identify the diseases timely.Africa is an underdeveloped region,and it is objectively difficult to identify the diseases by technicians in the fields or expert remote consultation.A cheap and efficient solution is to use deep learning algorithms to classify photos of cassava leaf taken by farmers in real time,but need to train a high-accuracy classifier based on the given dataset.This thesis selects the cassava leaf dataset from Uganda.Exploratory data analysis was performed to understand the quantity,quality and distribution of the data,and also to explore channel information of images in different categories.It shows that the dataset has some problems such as imbalanced class,poor image quality and noise.Therefore,stratified 5-fold cross validation and data enhancement techniques were used to preprocess,aiming to improve the robustness of the model and enhance the identifiability of image features.These operations lay the foundation for improving model performance.The dataset is successively used for five deep learning models including VGG16,Inception V3,Xception,Res Next50 and EfficientNetB3,to experiment based on transfer learning.By discussing the iteration situation and local test result of models,considering the overall and category evaluation indices,EfficientNetB3 lightweight model with better comprehensive performance is selected as the baseline model.The baseline model is further optimized while keeping fewer parameters.By introducing dropout,Bi-Tempered logistic loss,label smoothing and test time augmentation,the optimized EfficientNetB3 solved low accuracy problem caused by overconfidence,overcome the noisy influence in dataset on the decision boundary,and improved the generalization ability of the model.It is even superior to the others such as VGG16,Res Next50 with many times more parameters and FLOPs.Experiments show that the optimized EfficientNetB3 model utilized the attention mechanism to capture the boundary of the leaves’ disease region in the image more accurately,with less computation,higher accuracy and faster training speed.It increased the test accuracy from 87.16% of the baseline model to 88.79%.In practical applications,a higher classification success rate can be obtained by taking multiple photos of diseased plants with the rule of ‘less obey most’.