Page 183 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

172 Chapter 6 Plant leaf disease classification based on feature selection

Table 6.7 Confusion matrix of VGG16 with transfer learning.

Class C1 C2 C3 C4
C1 26 0 0 2
C2 4 8 0 4
C3 0 0 28 1
C4 6 2 0 4
With transfer learning, VGG16 also obtains better results. The training accuracy and
testing accuracy improve to 84.5% and 77.6% over the old results of 78% and 76%.

Table 6.8 Confusion matrix of ResNet-50 with transfer
learning.

Class C1 C2 C3 C4
C1 28 0 0 0
C2 6 4 0 3
C3 0 0 35 0
C4 4 0 0 5
With transfer learning, ResNet-50 achieved 84.71% testing accuracy and 85.39% of
training accuracy, which is a significant improvement in comparison with the former
results. Also, the model is much more stable.

4.3 Multilayer perceptron approach
4.3.1 Feature extraction
Contrast-limited adaptive histogram equalization (CLAHE)
was proposed by K. Zuiderveld in 1994 [17]. The method exam-
ines a histogram of intensities in a contextual region centered at
each pixel and sets the displayed intensity at the pixel as the
rank of that pixel’s intensity in its histogram. That histogram is a
modified form of the ordinary histogram in which the contrast
enhancement induced by the method at each intensity level is
limited to a user-selectable maximum. In this study, CLAHE is uti-
lized to perform thresholding of image. The original image is con-
verted to HSV format, and CLAHE is applied to the H channel to
enhance the contrast of the defective regions. The defective re-
gions are then separated and mapped back to the original image.