Context-CAM: Context-Level Weight-Based CAM With Sequential Denoising to Generate High-Quality Class Activation Maps.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Class activation mapping (CAM) methods interpret deep convolutional neural network (CNN) decisions and aid weakly supervised semantic segmentation (WSSS).
• Existing CAM methods struggle with complete object coverage and often include background noise in generated maps.

Purpose of the Study:

• To introduce an innovative Context-level weights-based CAM (Context-CAM) method.
• To address limitations of traditional CAM methods regarding object coverage and background noise.

Main Methods:

• Developed a Region-Enhanced Mapping (REM) module utilizing context-level weights to highlight non-discriminative yet relevant regions.
• Implemented a Semantic-guided Reverse Sequence Fusion (SRSF) strategy for sequential denoising and fusion of enhanced maps from deep layers to shallow layers.

Main Results:

• Context-CAM significantly improves class activation map quality, outperforming existing methods on the Energy-Based Pointing Game (EBPG) score by up to 35.49%.
• The method effectively enhances object coverage and reduces background noise compared to state-of-the-art approaches.
• Context-CAM seamlessly integrates into existing WSSS frameworks, boosting segmentation performance without architectural changes.

Conclusions:

• Context-CAM offers a superior approach to generating class activation maps, enhancing interpretability and segmentation accuracy.
• The proposed REM and SRSF modules provide effective solutions for common CAM limitations.
• This method holds significant potential for advancing WSSS tasks and deep learning model interpretability.