Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MicroCrackAttentionNeXt: Advancing Microcrack Detection in Wave Field Analysis Using Deep Neural Networks through Feature Visualization

Fatahlla Moreh, Yusuf Hasan, Bilal Zahid Hussain, Mohammad Ammar, Sven Tomforde

TL;DR

This study proposes an asymmetric encoder–decoder network with an adaptive feature reuse block for microcrack detection using DNNs through an automated pipeline using wave fields interacting with the damaged areas.

Abstract

Micro Crack detection using deep neural networks (DNNs) through an automated pipeline using wave fields interacting with the damaged areas is highly sought after. These high-dimensional spatio-temporal crack data are limited, and these datasets have large dimensions in the temporal domain. The dataset presents a substantial class imbalance, with crack pixels constituting an average of only 5% of the total pixels per sample. This extreme class imbalance poses a challenge for deep learning models with the different micro-scale cracks, as the network can be biased toward predicting the majority class, generally leading to poor detection accuracy. This study builds upon the previous benchmark SpAsE-Net, an asymmetric encoder-decoder network for micro-crack detection. The impact of various activation and loss functions were examined through feature space visualization using the manifold discovery and analysis (MDA) algorithm. The optimized architecture and training methodology achieved an accuracy of 86.85%.

MicroCrackAttentionNeXt: Advancing Microcrack Detection in Wave Field Analysis Using Deep Neural Networks through Feature Visualization

TL;DR

This study proposes an asymmetric encoder–decoder network with an adaptive feature reuse block for microcrack detection using DNNs through an automated pipeline using wave fields interacting with the damaged areas.

Abstract

Micro Crack detection using deep neural networks (DNNs) through an automated pipeline using wave fields interacting with the damaged areas is highly sought after. These high-dimensional spatio-temporal crack data are limited, and these datasets have large dimensions in the temporal domain. The dataset presents a substantial class imbalance, with crack pixels constituting an average of only 5% of the total pixels per sample. This extreme class imbalance poses a challenge for deep learning models with the different micro-scale cracks, as the network can be biased toward predicting the majority class, generally leading to poor detection accuracy. This study builds upon the previous benchmark SpAsE-Net, an asymmetric encoder-decoder network for micro-crack detection. The impact of various activation and loss functions were examined through feature space visualization using the manifold discovery and analysis (MDA) algorithm. The optimized architecture and training methodology achieved an accuracy of 86.85%.

Paper Structure

This paper contains 14 sections, 2 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. RELATED WORK
  3. METHOD
  4. Wave field data
  5. MicroCracksAttentionNeXt Model Architecture
  6. Training Procedure
  7. Activation Functions
  8. Loss Functions
  9. Metrics
  10. Results & Discussion
  11. MDA Analysis
  12. Conclusion and Future work
  13. Conclusion
  14. Future work

Figures (5)

  • Figure 1: 1) The 6 frames (100 time-steps interval, from left to right) of a displacement() wave propagation inside the defined plate with cracksWuttke2021. 2) Visualization of a data instance.
  • Figure 2: MicroCrackAttentionNeXt model architecture.
  • Figure 3: 1) MDA visualizations of layers using Gelu activation and Dice loss, shown for a) Layer 22, b) Layer 25, c) Layer 34, and d) Layer 64 and, 2) MDA visualization of Layer 64 utilizing different activation functions: a) ELU, b) ReLU, c) GELU, and d) SELU.
  • Figure 4: MDA visualization of Layer 64 comparing a) Untrained Model and b) Trained Model.
  • Figure 5: MDA visualization comparing a) 1D-Densenet Moreh2024 and b) Our proposed model - MicroCrackAttentionNeXt. The highlighted region in black indicates the region where the cluster is broken in 1D-Densenet. In contrast, the same region in MicroCrackAttentionNeXt shows coherency implying that the MicroCrackAttentionNeXt learned good feature representations for microcracks.