The Effect of Annotation Quality on Wear Semantic Segmentation by CNN

Area of Science:

• Materials Science
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are increasingly used for analyzing material wear.
• Accurate semantic segmentation of wear is crucial for predicting tool life and performance.
• Titanium nitride (TiN) and titanium carbonitride (TiCN) coatings are vital in cutting tools.

Purpose of the Study:

• To evaluate the influence of annotation quality and domain expertise on CNN performance for wear segmentation.
• To compare model performance on TiN and TiCN coated end mills.
• To identify factors affecting segmentation accuracy and model robustness.

Main Methods:

• Development of an innovative measurement system for wear analysis.
• Customized CNN architecture for semantic segmentation of wear.
• Comparison of model performance across annotators with varying domain expertise.
• Analysis of model sensitivity to annotation inconsistencies and hyperparameters.

Main Results:

• Domain expertise significantly impacted CNN performance, with higher expertise yielding superior mean Intersection over Union (mIoU) scores.
• Annotator 1 (high expertise) achieved mIoU of 0.8153 (abnormal wear) and 0.7120 (normal wear) on TiN datasets.
• Models for TiCN datasets exhibited higher coefficient of variation (16.32%) compared to TiN (8.6%) due to subtle wear characteristics.

Conclusions:

• Annotation quality and domain expertise are critical factors for successful wear segmentation using CNNs.
• Subtle wear features in TiCN coatings present challenges for accurate segmentation.
• Optimized annotation policies and high-quality imaging are essential for improving wear segmentation models.