Week6 — Plant Disease Detection
Theme: Classification of plants that healthy or diseased and find photographed plant disease.
Team Members: Fatmanur Turhan, Sevda Sayan, İsmet Seyhan
Last week, we achieved to train our ResNet50 model and published our train/validation losses and accuracies in addition to VGG16 results published on the fourth week.
In the end of our experiments on training/validation datasets, ResNet50 model training by using the transfer learning method was chosen. This week we evaluated our model on unseen pictures and repeat experiments on our test data.
Obtained results are shown below.
As seen above, the best results are achieved on ResNet50 model that is trained using transfer learning method. Resnets are a kind of CNNs called Residual Networks. They are very deep compared to VGG, and Resnet 50 refers to a 50 layers Resnet. Resnet introduced residual connections between layers, meaning that the output of a layer is a convolution of its input plus its input. Moreover, layers in a Resnet also use Batch Normalization, which has also been incorporated to VGG.
Lets dive into our best model and visualize weights. As suspected, they all look like somewhat blurry versions of our diseased or healthy plant leaves. They appear almost as though we averaged many images belonging to each class.
Suppose we want to detect an input from an image of a Tomato Early Blight. Neuron responsible for classifying Tomato Early Blights should have a high value because its weights are such that high weights tend to align with pixels tending to be high in Tomato Early Blights. We can comment these weights as forming templates of the output classes. This is fantastic we are not saying our network anything about what these leafs are, yet they remind us those classeses anyway.
Each of taught plant disease type have nearly 50 samples. The graph that shows our class based accuracies and confusion matrix is shown below.
When we focus on those visual results we can clearly say that model can classify them properly even on unseen data. Fifty percent of our classes have classified with accuracy of %100. Besides this, just %71.2 of Potato Late Blight classified correctly which has the lowest accuracy.
This was the end of our project and last blog.
Hope to see you in another projects…
