*4.2. Automatic Digit Recognition Using Deep Transfer Learning*

For Arabic digit recognition in the historical registers, we first created a small dataset of 50 digits obtained from the spotted one-digit numerals. The dataset is balanced in terms of digit types. We then used two large datasets for using deep transfer learning. The first one is HODA dataset [36] which includes 70,000 images. The second one is AHDBase [37] which is also composed of 70,000 digits written by 700 participants. Lastly, we presented results obtained by dividing our local dataset (see Figure 7 for samples) into 80% training and 20% test sets.

**Figure 7.** All ten-digit samples from our local data. Fifty digits are selected from the output of the numeral spotting system.
