Table 4.
Grad-CAM visualization of correctly classified subjects from the AMI dataset. The model looks at various places around the ear structure, exploiting these auxiliary parts to make predictions. This is justifying the drop in recognition rate on the AMIC dataset where these extra parts are removed from the profile images.
Dataset | Original | Grad-CAM | Heatmap | Original | Grad-CAM | Heatmap |
---|---|---|---|---|---|---|
AMI |
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |