Table 3.

Machine identification methods.

CNN model	Description
ResNet	Residual networks make use of residual blocks to train very deep neural networks. Making use of these blocks helps prevent the complications that come with training deep networks such as accuracy degradation. The key idea behind these blocks is to have the output of one layer be sent as an input to a layer deeper in the network. This is then added with layer's output in its normal path before the nonlinear function is applied. The variants of ResNet include ResNet-50, ResNet-1202, ResNet-110, ResNet-164, ResNet-101, and ResNet-152 [30].
DenseNet	In the DenseNet architecture, for any layer in the convolutional neural network, the input will be the concatenated result of the output of all the previous layers in the network. By combining information from the previous layers, the computational power can be reduced. Through these connections across all layers, the model is able to retain more information moving across the network improving the overall performance of the network. The variants of DenseNet are DenseNet-B and DenseNet-BC [31].
MobileNet	MobileNet is an architecture that is implemented for mobile devices. They use depth-wise separable convolutions where the filters are added for individual input channels as opposed to the general CNN models. These each channel separate and stack the convoluted outputs together. The different variants include MobileNetV1 and MobileNetV2 [32].
ShuffleNet	They are efficient convolutional neural networks which use filter shuffling to reduce computational complexity. They can reduce the time required to train a model and are used most commonly in smartphones. Shuffling the filters allows more information to pass the network which increases the accuracy of the network with a limited number of resources. The different ShuffleNet models include ShuffleNetV1 and ShuffleNetV2 [33].