A Self-Supervised Model Based on CutPaste-Mix for Ductile Cast Iron Pipe Surface Defect Classification.

Area of Science:

• Industrial automation and machine vision.
• Materials science and non-destructive testing.

Background:

• Manual inspection of industrial components like ductile cast iron pipes (DCIP) generates vast datasets, making defect identification costly and time-consuming.
• Existing automated systems often require extensive labeled data for training, posing a challenge for defect classification.

Purpose of the Study:

• To develop a cost-effective, self-supervised binary classification algorithm for automated defect detection in DCIP images.
• To reduce the manual effort and cost associated with identifying defect images in industrial settings.

Main Methods:

• A self-supervised binary classification algorithm was employed for defect image classification.
• The CutPaste-Mix data augmentation strategy was utilized to enhance defect-free and defect data.
• A deep convolutional neural network was combined with Gaussian Density Estimation to compute anomaly scores for classifying abnormal regions.

Main Results:

• The proposed method achieved robust performance on both a dedicated DCIP image dataset and in practical field applications.
• An impressive Area Under Curve (AUC) of 99.5 was attained, demonstrating high classification accuracy.
• The system proved to be a cost-effective solution for data support in subsequent DCIP surface inspection model training.

Conclusions:

• The self-supervised approach offers an efficient and accurate method for defect detection in industrial settings.
• This technique significantly reduces the cost and manual effort in identifying defects in ductile cast iron pipes.
• The developed algorithm provides valuable data support for training advanced surface inspection models.