**4.3 CNN architecture and benchmark dataset**

The network structure used for this study is made of the following layers: four convolutions, three normalizations, one ReLU, one softmax, one classification, and two pooling. The highlighted blocks indicated in **Figure 7** are the pooling layers. Each pooling method has been implemented in those layers, utilizing one single method for the entire network. All methods have been tested for the dataset considering a 2*x*2 region, 2*x*2 stride and no padding, except for the Lift. Random Channel which uses the entire image channel as a region.

The benchmark dataset used for training and testing is the MNIST, as shown in **Figure 8**. The MNIST dataset [42] contains a large number of color images of handwritten single digits divided into 10 classes, and their size is 28*x*28 pixels. The training set is made of 60,000 images and the testing set of 10,000 images.
