4.2.2. CNN-Based Handwritten Arabic Digit Recognition on the Local Data

We separated the local data into 80% training and 20% test sets. The local dataset is balanced in terms of digit types. Then, we applied the CNN architecture shown in Figure 8 to train a model for the local data and tested on the remaining separate test set.

**Figure 8.** The CNN architecture is shown. The layers in the green rectangle are used for feature extraction in DTL. In order to train a model for local data, all layers are used. Conv2D stands for 2D

Convolutional Layer.
