Automated Industrial Composite Fiber Orientation Inspection Using Attention-Based Normalized Deep Hough Network

Area of Science:

• Materials Science
• Computer Vision
• Image Processing

Background:

• Fiber-reinforced composites (FRC) are crucial materials with properties dependent on fiber orientation.
• Automated visual inspection using image processing is vital for measuring fiber orientation in FRC.
• Existing methods like the deep Hough Transform (DHT) struggle with background and long-line anomalies.

Purpose of the Study:

• To develop an improved method for accurate fiber orientation measurement in FRC.
• To address the limitations of the deep Hough Transform (DHT) in handling real-world image anomalies.
• To enhance the reliability of automated visual inspection for FRC materials.

Main Methods:

• Introduction of deep Hough normalization to mitigate sensitivity to line segment anomalies.
• Development of an attention-based deep Hough network (DHN) to eliminate background anomalies.
• Extensive evaluation on three established datasets simulating real-world anomaly scenarios.

Main Results:

• The proposed deep Hough normalization effectively detects short, true "line-like" structures.
• The attention-based DHN successfully eliminates background noise and identifies key fiber regions.
• The DHN achieves competitive performance in F-measure, Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE) compared to state-of-the-art methods.

Conclusions:

• The developed attention-based deep Hough network (DHN) offers a robust solution for fiber orientation measurement in FRC.
• The method demonstrates significant improvements in handling image anomalies common in industrial applications.
• This advancement contributes to more reliable automated quality control for fiber-reinforced composites.