Week4 — Plant Disease Detection

Theme: Classification to plants that healthy or diseased and predict to photographed plant disease.

Team Members: Sevda Sayan, Fatmanur Turhan, İsmet Seyhan

In our blog last week, we explained our approach and libraries we will use. In this blog, we will explain experiments that we are working on and indicate some initial results.

We have decided to use the usefulness of deep convolutional neural networks for our classification. Up to this time we focused on VGG16 architecture proposed by K. Simonyan and A. Zisserman from the University of Oxford. They submitted it to ILSVRC-2014 and became one of the famous model. It is a pretty large network and it has approximately 138 million parameters.

Experimental configurations that we will focus on by changing the following parameters are indicated below

Configurations About Architecture

1. VGG16
2. Resnet50

Configurations About Training mechanism

1. Transfer Learning
2. Training from Scratch

As a beginning, we have trained our VGG16 model from scratch by using our belittled dataset. But gotten losses and accuracies was not satisfying. As seen graphs below, train accuracy start with 4% and end with 16% when validation accuracy start with 10% and end with 33%.

Then we have trained VGG16 using transfer learning on our dataset. As a result we obtained higher scores on both train and validation phases. Train accuracy start with 54% and end with 85% when validation accuracy start with 81% and end with 94%.

In this experiments we used Adam optimizer and learning rate is setted to 0.001. Each experiments runs for a total of 20 epochs, where epoch refers to one cycle through the full training dataset. On each epoch we usually observed decreasing losses and increasing accuracies as it should be.

Among the transfer learning and training from scratch mechanisms, transfer learning consistently performs better than training from scratch and always yields better results which were expected.

After this point, we will complete our experiments and choose the best model. We are estimating that a model composed of Resnet50 by using transfer learning method will fit to our problem. After that we’ll observe class based accuracies and try to improve insufficient ones.

Next week we will indicate our next experiments that about Resnet50.

Thank you for reading…