Optimized Dropkey-Based Grad-CAM: Toward Accurate Image Feature Localization.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Gradient-weighted Class Activation Mapping (Grad-CAM) is a key technique for feature localization in image recognition, providing insights into neural network decision-making.
• Standard Grad-CAM-based deep convolutional neural network (CNN) models demonstrate limitations in resisting large-scale noise interference.

Purpose of the Study:

• To optimize deep CNN models for enhanced noise resistance and accurate feature localization.
• To evaluate the effectiveness of the Dropkey algorithm in improving Grad-CAM's performance against noise.

Main Methods:

• An optimized deep CNN model was developed by integrating the Dropkey algorithm with Grad-CAM.
• The improved Grad-CAM applies an attention mechanism to the feature map before gradient calculation, introducing randomness and masking attention scores.
• Dropout was used as a comparative method to assess the Dropkey algorithm's efficacy.

Main Results:

• The Dropkey-enhanced Grad-CAM deep CNN model significantly improved resistance to large-scale noise interference.
• Under noise variance of 0.6, the Dropkey-enhanced ResNet50 model achieved a prediction confidence of 0.878, outperforming other models.
• The optimized model demonstrated robust performance in visualizing features affected by distortion, low contrast, and small object characteristics.

Conclusions:

• The Dropkey algorithm effectively enhances Grad-CAM's robustness against noise in deep CNNs.
• This optimized approach achieves accurate feature localization and visualization, even under challenging image conditions.
• The enhanced Grad-CAM shows significant potential for practical computer vision applications, including autonomous driving, for verifying model understanding of environmental elements.