A dataset for surface defect detection on complex structured parts based on photometric stereo

Area of Science:

• Computer Vision
• Machine Learning
• Materials Science

Background:

• Automated Optical Inspection (AOI) is vital for industrial quality control.
• Traditional AOI faces challenges with shadows, reflectivity, and non-planar surfaces, leading to detection inaccuracies.

Purpose of the Study:

• To develop a novel defect detection technique for metal surfaces overcoming AOI limitations.
• To create a comprehensive dataset for training and validating deep learning models for metal surface defect detection.

Main Methods:

• Proposed a Stroboscopic Illuminant Image Acquisition (SIIA) method combining photometric stereo vision and deep learning.
• Developed a Taylor Series Channel Mixer (TSCM) to create pseudo-color images from multi-angle illuminations.
• Utilized hue randomization for data augmentation and validated object detection models (FCOS, YOLOv5, YOLOv8, RT-DETR) on the Metal Surface Defect Dataset (MSDD).

Main Results:

• Achieved a mean Average Precision (mAP) of 86.1% on the MSDD, outperforming traditional methods.
• The proposed technique effectively handles shadows and reflectivity issues inherent in AOI.
• The MSDD comprises 138,585 single-channel and 9,239 mixed images covering eight defect types.

Conclusions:

• The novel deep learning and photometric stereo vision approach significantly enhances automated visual inspection of metal surfaces.
• The developed MSDD provides a valuable resource for advancing research in industrial defect detection.
• The method offers a robust solution for end-to-end defect detection using universal object detectors.