Page 174 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

Chapter 6 Plant leaf disease classification based on feature selection 163

Figure 6.7 AlexNet architecture.

The convolutional layers are split to contain mapped kernels
in the same graphical processing units. The convolutional layers
reduce the image parameters producing 4096 dimensional
features for the FC layer that are mapped into a logistic regression
output containing 1000 classes. This architecture uses various
kinds of transformations for data augmentation to increase
learning. These spatial transforms provide more robust training
samples for the network.
3.3.2 VGG16 (2014)
VGGNet [15] is a neural network that performed extremely well
in the ImageNet Large Scale Visual Recognition Challenge
(ILSVRC) in 2014. This architecture is from VGG group, Oxford.
It makes the enhancement over AlexNet by replacing expensive
kernel size filters with various 33 filters in a steady progression.
Fig. 6.8 shows the VGG16 architecture.
This idea of blocks/modules was introduced in VGG. The
VGG convolutional layers are trailed by three FC layers. The
width of the filters begins with 64 and increments by a factor
of 2 after each pooling layer. The images in the challenge were
divided into 1000 unique classes. Given a test image, the VGG
network determines a likelihooddin the range of 0 and 1dfor
every one of those 1000 classes and selects the class with the
highest likelihood.