A Dynamic Weights-Based Wavelet Attention Neural Network for Defect Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Industrial defect detection is crucial for production quality.
• Current deep learning methods struggle with weak defects and strong background noise.
• Enhanced feature representation and noise reduction are needed for improved accuracy.

Purpose of the Study:

• To propose a novel deep learning model, the dynamic weights-based wavelet attention neural network (DWWA-Net), for enhanced industrial defect detection.
• To address limitations in detecting weak defects and defects under significant background noise.
• To improve the overall accuracy and robustness of automatic defect detection systems.

Main Methods:

• Development of a dynamic weights-based wavelet attention neural network (DWWA-Net).
• Integration of wavelet neural networks and dynamic wavelet convolution networks (DWCNets) for noise filtering and model convergence.
• Implementation of a multiview attention module to focus on potential defect targets.
• Inclusion of a feature feedback module to enhance defect feature information.

Main Results:

• The proposed DWWA-Net demonstrates superior performance compared to state-of-the-art methods.
• Significant improvements in mean precision were observed on benchmark datasets (GC10-DET: 6.0% increase, NEU: 4.3% increase).
• The method effectively enhances feature representation and reduces background noise.

Conclusions:

• The DWWA-Net provides a robust solution for challenging industrial defect detection scenarios.
• The model's ability to handle weak defects and noisy images marks a significant advancement.
• DWWA-Net shows potential for broad application across various industrial inspection fields.