Page 178 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

Chapter 6 Plant leaf disease classification based on feature selection 167

of disease. This gives an advantage that a huge amount of data is
used to train this model, which is able to learn features efficiently.
Thus, based on PlantVillage data set, a pretrained model is accom-
plished, which is used for transfer learning to mango disease data
set. Besides, for mango disease data set, we separate randomly the
entire data set to two subsets: training and test set. The training
was done by the training set, whereas the evaluation is carried
on the test set. The cross-entropy is applied as a loss function to
estimate the error prediction after classification layer. Moreover,
the optimization algorithm for the training process is done by
stochastic gradient descent with momentum (SGDM) to update
weights and biases at each iteration, which only a requires small
number of epochs for training. The maximum number of epochs
to train the proposed model is 30 with an initial learning rate
0.0005 and dropping every 10 epochs by 1/10.

4. Results

Our data are split into training and test sets, with 80% of the
data being the training and 20% being the test data. Then the
training set is split again to 80% for training and 20% for
validation.
All the models are implemented on a Desktop PC having GPU
GTX 1070, which has 1920 CUDA cores with processor Intel(R)
Core(TM) i7-7700 at 3.6 GHz, 32 GB of DDR4 random access
memory (RAM), and a solid-state drive (SSD) of 128 GB
(Tables 6.1 and 6.2).

Table 6.1 Confusion matrix of VGG model.

Class C1 C2 C3 C4
C1 34 3 0 0
C2 7 5 0 0
C3 3 0 24 0
C4 5 0 0 4

C1, anthracnose; C2, gall midge; C3, healthy; C4, powdery mildew.