We develop a novel post-hoc visual explanation method called Augmented Score-CAM that improves Score-CAM activation maps by applying image augmentation based on the matrix computation techniques [1], Score-CAM [2], and Augmented Grad-CAM [3]. In our experiments, we denote our model as ASC and Score-CAM as SSC.
The following is the pipeline of our method:
Paper: Augmented Score-CAM: High resolution visual interpretations for deep neural networks published at: Elsevier Knowledge Based Systems
You can run an example for Augmented Score-CAM via Google Colab
If you find the code useful for your research, please cite our work:
@article{IBRAHIM2022109287, title = {Augmented Score-CAM: High resolution visual interpretations for deep neural networks}, journal = {Knowledge-Based Systems}, volume = {252}, pages = {109287}, year = {2022}, issn = {0950-7051}, doi = {https://doi.org/10.1016/j.knosys.2022.109287}, url = {https://www.sciencedirect.com/science/article/pii/S0950705122006451}, author = {Rami Ibrahim and M. Omair Shafiq}, keywords = {Explainable AI, Class activation maps, Augmented Score-CAM}
If you have any questions, feel free to contact me at: ramif.ibrahim@carleton.ca
[1] AIR Galarza, J Seade (2007). Introduction to classical geometries. Springer Science & Business Media.
[2] Wang, Haofan, et al. (2020). Score-CAM: Score-weighted visual explanations for convolutional neural networks. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops (pp. 24-25).
[3] Morbidelli, Pietro, et al. (2020). Augmented Grad-CAM: Heat-maps super resolution through augmentation. ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 4067-4071).