Page 173 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

162 Chapter 6 Plant leaf disease classification based on feature selection

24@48x48
24@16x16
8@128x128 1x256
8@64x64
1x128
1x4

Max-Pool Convolution Max-Pool Dense
Figure 6.6 An example of a CNN.

the other. The preprocessing required in a ConvNet is much
lower as compared with other classification algorithms. While
in primitive methods, filters are hand-engineered, with enough
training, ConvNets can learn these filters or characteristics
(Fig. 6.6).
In this study, we explore three popular CNN architectures,
namely, AlexNet, VGG16, and ResNet to build the classification
system.
3.3.1 AlexNet (2012)
AlexNet [14] was an innovative architecture to tackle large
labeled data sets for image recognition with higher precision
and efficiency. AlexNet introduced various key design innovations
like the addition of dropout layers for higher accuracy along with
incorporation of distributed processing for better scalability and
faster training. Hence, leveraging a multi-GPU system can speed
up training as well as evaluation for large data sets. AlexNet sets
the premise to better image classification architectures and
research using DL techniques and methods. Fig. 6.7 shows the
architecture of AlexNet.
The architecture for the AlexNet contains five convolution
neural layers and three fully connected (FC) layers. The architec-
ture of AlexNet used ReLU activation function in its neural layers
to accommodate faster training over traditional activation func-
tions like tanh. The response to normalization of layer’s average
data given during training to prevent stagnant learning
iterations and high false positives in recognition. The max pool-
ing layers helped reduce variance also capturing strong inputs
over the network layers. Pooling layers are placed after the
response normalization layers. The architecture uses overlapping
pooling in its structure.