Table 4.

The advantages and disadvantages of the AlexNet, ResNet, and GoogLeNet CNN architecture [34,35].

Architecture	Advantage	Disadvantage
AlexNet layer depth: 8 parameters: 60 million	There is low feature loss, as the ReLU activation function does not limit the output. Uses data enhancement, dropout, and normalization layers to prevent the network from overfitting and improve the model generalization.	This model has much less depth; hence, it struggles to learn features from image sets. Takes more time to achieve higher accuracy (highest accuracy achieved: 99.11%).
ResNet-50 layer depth: 50 parameters: 25.6 million ResNet-101 layer depth: 101 parameters: 44.5 million	Decreased the error rate for deeper networks by introducing the idea of residual learning. Instead of widening the network, the increased depth of the network results in fewer additional parameters. This greatly reduces the training time and improves accuracy (highest accuracy ResNet-50: 99.20%; highest accuracy ResNet-101: 99.01%). Mitigates the effect of vanishing gradient.	A complex architecture Many layers may provide very little or no information. Redundant feature-maps may happen to be relearned.
GoogLeNet layer depth: 22 parameters: 7 million	Computational and memory efficiency. Reduced number of parameters by using bottleneck and global average pooling layer. Use of auxiliary classifiers to improve the convergence rate.	Lower accuracy (highest accuracy: 98.77%). Its heterogeneous topology necessitates adaptation from module to module. Substantially reduces the feature space because of the representation bottleneck and thus sometimes may lead to loss of useful information.