Page 90 - Read Online
P. 90

Bah et al. Intell Robot 2022;2(1):72­88  I http://dx.doi.org/10.20517/ir.2021.16          Page 84
































                   Figure 7. The accuracy and the loss of the training and validation data of FERGIT on the ResNet based model over 100 epochs.
































                       Figure 8. The accuracy and the loss of the training and validation data of CK+ on the basic model over 100 epochs.


               CK+, 97%. See Table 4. These results express the superiority of the presented methodology compared to best
               results with CNN architectures such as Pu [19]  who achieved an accuracy of 95.74%, and Cheng [44]  achieved
               success rate of 94.4%. See Table 5.




               4. DISCUSSION
               Our CNN FER model, which is based on ResNet, took 48 minutes to learn multiple facial images and then
               distinguish between seven (7) emotions, although the number of parameters is relatively big (9,766,391 pa-
               rameters). The traditional CNN on the other hand took 44 minutes to run 100 epochs, but without improving
   85   86   87   88   89   90   91   92   93   94   95