Transformation-invariance properties of Grad-CAM in Convolutional Neural Networks

The widespread adoption of Convolutional Neural Networks (CNNs) in image classification tasks has led to increasing interest in interpreting their decisions. Gradient weighted Class Activation Mapping (Grad-CAM) is a post-hoc explanation technique that generates visual heatmaps to highlight the class-relevant regions in an input image. However, in real-world scenarios, input images often undergo transformations such as rotation, zoom, and shifts along the horizontal and vertical axis, which may compromise the reliability of Grad-CAM explanations. This research systematically investigates the transformation-invariance properties of Grad-CAM across popular CNN architectures—ResNet152, DenseNet201, and Xception. An experimental pipeline was designed and implemented to apply controlled transformations to input images, generate corresponding Grad-CAM heatmaps, and quantitatively assess their consistency using metrics like the Euclidean (L2) difference and AUC-ROC analysis. Through extensive experiments, we demonstrate that Grad-CAM heatmaps exhibit varying degrees of inconsistency under transformations. To mitigate this variability, we propose and validate heatmap averaging methods, producing robust, transformation invariant heatmaps. This investigation provides insights into Grad-CAM’s robustness limitations and presents techniques to enhance the reliability of visual explanations in CNN-based image classification tasks.

Name thesis:degree_name

Master of Science (MSc)

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Ontario Institute of Technology

Year dc:date.issued

2025

Author dc:creator

• Roy, Emon

Advisors dc:contributor.advisor

• Ebrahimi, Mehran
• Davoudi, Kourosh

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10155/2057

OAI identifier oai:identifier

oai:ontariotechu.scholaris.ca:10155/2057