SSG-CAM: enhancing visual interpretability through refined second-order gradients and evolutionary multi-layer fusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Class Activation Mapping (CAM) is crucial for understanding Convolutional Neural Networks (CNNs).
• Existing first-order gradient methods suffer from noise and gradient saturation, limiting localization accuracy.
• Manual feature map selection hinders the integration of multi-level information, weakening visual explanations.

Purpose of the Study:

• To develop an advanced CAM technique that overcomes the limitations of current methods.
• To enhance the accuracy and robustness of visual explanations derived from CNNs.
• To automate the selection and fusion of optimal feature maps for improved interpretability.

Main Methods:

• Proposed Smooth Second-Order Gradient Class Activation Mapping (SSG-CAM) utilizing second-order gradients to mitigate saturation and noise.
• Integrated SSG-CAM with the Differential Evolution (DE) algorithm to form the DE-SSG-CAM framework for automated multi-level feature map optimization.
• Employed a smoothing technique to reduce noise in gradient calculations.

Main Results:

• SSG-CAM effectively alleviates gradient saturation and reduces noise compared to first-order methods.
• The DE-SSG-CAM framework demonstrated superior performance in weakly supervised object localization and semantic segmentation tasks.
• Achieved a mean Intersection over Union (mIoU) of 62.38% for localizing malarial parasite lesions, showcasing high accuracy in biomedical image analysis.

Conclusions:

• SSG-CAM provides a more accurate and robust visual explanation tool for CNNs.
• The DE-SSG-CAM framework offers an automated approach to distilling optimal visual interpretations from deep neural networks.
• This work advances the field of explainable AI, particularly in complex biomedical imaging applications.