Figure 1
From: Insights and approaches using deep learning to classify wildlife
Comparison between Grad-CAM, GBP, and GG-CAM. Once trained, any image (leftmost panel) can be overlaid with its Grad-CAM heat map (left center panel) to identify the region of ‘most interest’ to the CNN (see Appendix 4). Similarly, the corresponding feature map (center right panel), produced using Guided Back-propagation (GBP), (which, as described in Appendix 4, identifies the most important visual features to our CNN) can be weighted by the Grad-CAM heat map to produce the guided Grad-CAM (GG-CAM) image seen in the rightmost panel. Note that in this Nyala image, GBP is less discriminative than GG-CAM: both highlight the stripes of the Nyala, whereas GPB includes non-species-discriminative tree branches and legs.
