Deterioration Level Estimation Based on Convolutional Neural Network Using Confidence-Aware Attention Mechanism for...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Structural Engineering

Background:

• Infrastructure inspection relies on accurate deterioration level estimation.
• Conventional spatial attention mechanisms in deep learning can be unreliable due to ineffective attention maps.
• This limitation hinders precise feature highlighting for accurate estimation.

Purpose of the Study:

• To develop an improved attention mechanism for convolutional neural networks (CNNs) in infrastructure inspection.
• To address the unreliability of conventional attention mechanisms by introducing confidence awareness.
• To enhance the accuracy of deterioration level estimation in infrastructure.

Main Methods:

• A novel confidence-aware attention mechanism is proposed for CNNs.
• Confidence is calculated using the entropy of estimated class probabilities during attention map generation.
• This mechanism reduces the impact of ineffective attention maps by incorporating confidence scores.

Main Results:

• The proposed confidence-aware attention mechanism effectively utilizes attention maps by considering their confidence.
• The method demonstrates improved focus on critical regions for final deterioration level estimation.
• Experimental results on real-world infrastructure inspection images validate performance improvements.

Conclusions:

• The confidence-aware attention mechanism significantly enhances CNN-based deterioration level estimation.
• This approach offers a more robust and accurate method for infrastructure inspection.
• The findings contribute to advancing automated defect detection and structural health monitoring.