HiRA-CAM: Preserving Fine-Grained Spatial Relevance in Gradient-Based Visual Explanations

Manasi Nerurkar; Ali A. Minai

cs.CVAug 19, 2026

HiRA-CAM: Preserving Fine-Grained Spatial Relevance in Gradient-Based Visual Explanations

Authors: Manasi Nerurkar, Ali A. Minai

Organizations: Dept. of Electrical and Computer Engineering University of Cincinnati Cincinnati, USA

Abstract

Deep Learning models can include billions of parameters or more, making it difficult to explain their internal transformations and outputs. However, explainability is increasing in importance due to the use of AI in crucial applications. This paper focuses on the interpretability of convolutional neural networks (CNNs). Building on the popular gradient based method LayerCAM for extracting internal features in CNNs, we propose an improved method named HiRA-CAM, and show that it outperforms both LayerCAM and Grad-CAM on creating useful saliency maps for object classification. The main feature of HiRA-CAM is its adaptive use of activation maps from all the layers of the CNN to arrive at a more focused saliency map.

Figures & tables

Backbone	Method	loc1 (%)	loc5 (%)	Pointing	Del. AUC
VGG16	Grad-CAM	41.8	52.6	60.2	0.44
	LayerCAM	44.3	55.1	64.8	0.38
	HiRA-CAM	45.9	56.7	66.5	0.34
ResNet-50	Grad-CAM	38.9	49.8	56.7	0.47
	LayerCAM	42.6	53.4	62.9	0.40
	HiRA-CAM	45.1	55.8	68.1	0.33

TABLE I : Comparison of localization and faithfulness metrics on ImageNet WSOL using ImageNet-pretrained CNN backbones.

Fig. 1 : Qualitative comparison of class activation maps generated using Grad-CAM, Layer-CAM, and the proposed HiRA-CAM on ImageNet validation images using VGG-16 (rows 1-2), ResNet-50 (rows 3-4) and DenseNet-121 (rows 5-6). HiRA-CAM produces more compact, and object-focused saliency maps, suppressing background activations.

Panel of Figure 3 — Fig. 2 : Bounding box localization from Grad-CAM, Layer-CAM, and HiRA-CAM on ImageNet using VGG-16 (top), ResNet-50 (Middle), and DenseNet-121 (bottom). The ground truth boxes are denoted by green and the red ones denote the predicted bounding boxes.

Fig. 3 : Deletion metric curves across different architectures. From top to bottom: VGG-16, ResNet-50, DenseNet-121. The data is averaged over 150 images in the validation set.

Explore similar work

Fusion-CAM: Integrating Gradient and Region-Based Class Activation Maps for Robust Visual Explanations
Mar 5, 2026Hajar Dekdegue, Moncef Garouani, Josiane Mothe +1Explainable Image Classification Feature Attribution
IRIT, UMR5505 CNRS Universit´e de Toulouse · IRIT, UMR5505 CNRS Universit´e Toulouse Capitole · INRAE, Centre Occitanie-Toulouse Unit´e Exp´erimentale APC
Class Activation Mapping in Explainable Computer Vision: A Method-Centered Review of CNN, Transformer, and Foundation-Model-Era Visual Explanations
Aug 12, 2026AmirHossein Eshghi, Hamid Saadatfar, Seyyed Ali Hoseini +2Transformer Interpretability Feature Attribution
Department of Computer Engineering, University of Birjand, Iran
DD-CAM: Minimal Sufficient Explanations for Vision Models Using Delta Debugging
Feb 22, 2026Krishna Khadka, Yu Lei, Raghu N. Kacker +1Explainable Image Classification Feature Attribution
Department of Computer Science & Engineering The University of Texas at Arlington Arlington, Texas, USA · Information Technology Laboratory National Institute of Standards and Technology (NIST) Gaithersburg, Maryland, USA

Mar 5, 2026J/K move · Enter open · S save

Fusion-CAM: Integrating Gradient and Region-Based Class Activation Maps for Robust Visual Explanations

Hajar Dekdegue, Moncef Garouani, Josiane Mothe, Jordan Bernigaud

IRIT, UMR5505 CNRS Universit´e de Toulouse · IRIT, UMR5505 CNRS Universit´e Toulouse Capitole · INRAE, Centre Occitanie-Toulouse Unit´e Exp´erimentale APC

Explainable Image Classification Feature Attribution Gradient-Based Attribution Saliency Map Evaluation Explainability Evaluation

Open paper AskPDF arXiv

Interpreting the decision-making process of deep convolutional neural networks remains a central challenge in achieving trustworthy and transparent artificial intelligence. Explainable AI (XAI) techniques, particularly Class Activation Map (CAM) methods, are widely adopted to visualize the input regions influencing model predictions. Gradient-based approaches (e.g. Grad-CAM) provide highly discriminative, fine-grained details by computing gradients of class activations but often yield noisy and incomplete maps that emphasize only the most salient regions rather than the complete objects. Region-based approaches (e.g. Score-CAM) aggregate information over larger areas, capturing broader object coverage at the cost of over-smoothing and reduced sensitivity to subtle features. We introduce Fusion-CAM, a novel framework that bridges this explanatory gap by unifying both paradigms through a dedicated fusion mechanism to produce robust and highly discriminative visual explanations. Our method first denoises gradient-based maps, yielding cleaner and more focused activations. It then combines the refined gradient map with region-based maps using contribution weights to enhance class coverage. Finally, we propose an adaptive similarity-based pixel-level fusion that evaluates the agreement between both paradigms and dynamically adjusts the fusion strength. This adaptive mechanism reinforces consistent activations while softly blending conflicting regions, resulting in richer, context-aware, and input-adaptive visual explanations. Extensive experiments on standard benchmarks show that Fusion-CAM consistently outperforms existing CAM variants in both qualitative visualization and quantitative evaluation, providing a robust and flexible tool for interpreting deep neural networks.

Abstract

Explore similar work

Fusion-CAM: Integrating Gradient and Region-Based Class Activation Maps for Robust Visual Explanations

Class Activation Mapping in Explainable Computer Vision: A Method-Centered Review of CNN, Transformer, and Foundation-Model-Era Visual Explanations

DD-CAM: Minimal Sufficient Explanations for Vision Models Using Delta Debugging