Page 182 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

Chapter 6 Plant leaf disease classification based on feature selection 171

Table 6.4 Confusion matrix of ResNet-50 model.

Class C1 C2 C3 C4
C1 34 0 0 1
C2 12 1 1 0
C3 4 0 22 0
C4 9 0 1 0
C1, anthracnose; C2, gall midge; C3, healthy; C4, powdery mildew. ResNet-50 model
achieved 67.06% testing accuracy and 63.46% training accuracy. We can see that
ResNet-50 provides better results than those of ResNet-18. The loss result is not very
good. Also the testing result was better than training result, indicating a high probability
of overfitting.

Table 6.5 Confusion matrix of AlexNet with transfer
learning.

Class C1 C2 C3 C4
C1 31 2 0 0
C2 5 15 0 0
C3 0 0 19 0
C4 5 4 2 2
AlexNet with transfer learning achieved 85.6% training accuracy and 78.8% testing
accuracy; as expected, it is a marked improvement over the conventional model (74.3%
training accuracy, 66% validation accuracy).

Table 6.6 Confusion matrix of ResNet-18 with transfer
learning.

Class C1 C2 C3 C4

C1 34 0 0 1
C2 6 13 0 0
C3 5 0 17 0
C4 5 0 0 4
Similar to AlexNet case, with transfer learning, we obtain better results with ResNet-18
model. The training accuracy and testing accuracy improve to 90.6% and 80% over the
old results of 84% and 68%.