Photo from Scientific Research

Visualization of saliency map from different methods. The saliency maps showcased in this figure stem from three inputs that were processed by a VGG-16 model trained on the ImageNet dataset. The Guided Grad-CAM, Guided Full-Grad, and Guided PANE generated by the Grad-CAM, Full-Grad, and PANE methods were multiplied by the saliency maps obtained through Guided BP, correspondingly. Compared with other methods, saliency maps from PANE focus more on discrete local features, i.e., certain pixels. From a statistical perspective, the saliency maps from PANE have a greater left-skewed distribution tendency.

DOI