MicrocrackAttentionNext: Advancing Microcrack Detection in Wave Field Analysis Using Deep Neural Networks Through...

Area of Science:

• Materials Science
• Computer Science
• Artificial Intelligence

Background:

• Automated microcrack detection using deep neural networks (DNNs) is crucial but hindered by limited, high-dimensional spatio-temporal data.
• Significant class imbalance in crack datasets, with crack pixels as low as 5%, challenges DNNs, leading to poor detection of microcracks.

Purpose of the Study:

• To develop an effective DNN model for microcrack detection that overcomes data limitations and class imbalance.
• To investigate the influence of different activation and loss functions on microcrack detection performance.

Main Methods:

• An asymmetric encoder-decoder network with an adaptive feature reuse block was proposed for microcrack detection.
• The manifold discovery and analysis (MDA) algorithm was employed for feature space visualization to analyze activation and loss functions.

Main Results:

• The proposed network architecture and training methodology demonstrated effectiveness in microcrack detection.
• An overall accuracy of 87.74% was achieved, indicating significant improvement in detecting microcracks.

Conclusions:

• The developed asymmetric encoder-decoder network with adaptive feature reuse is a promising approach for microcrack detection.
• The study highlights the importance of addressing class imbalance and optimizing network architecture for accurate microcrack identification.