A Reference Based Surface Defect Segmentation Algorithm For Automatic Optical Inspection System
Surface defect segmentation algorithms in Automatic Optical Inspection (AOI) system for modern manufacturing industries provide solutions to quality control with speed, volume and traceability. However, present complex algorithms which are accurate require high processing power using a large size of...
Saved in:
Main Author: | |
---|---|
Format: | Thesis |
Language: | English |
Published: |
2020
|
Subjects: | |
Online Access: | http://eprints.usm.my/51673/1/WONG%20ZE-HAO%20-%20TESIS%20cut.pdf |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Summary: | Surface defect segmentation algorithms in Automatic Optical Inspection (AOI) system for modern manufacturing industries provide solutions to quality control with speed, volume and traceability. However, present complex algorithms which are accurate require high processing power using a large size of learning dataset without labelling error. On the other hand, simple algorithms are not suitable for surfaces with complicated designs and variations. This study aims to develop an algorithm for the AOI system to segment and detect surface defects, requiring low processing power and a small number of learning dataset with labelling error resistance. Multiple Templates Anomaly Detection (MTAD) strategy is proposed to describe the local anomaly degree through distance functions computed from learning dataset. The learning dataset images are illumination normalized, registered and stacked across multiple templates in a kernel to form a histogram for each pixel location. Then, the histogram distance function for each location is computed using a pseudo-probability combination of novel histogram distance functions on a clustered histogram. Finally, surface defects are segmented from an anomaly heat map which is generated based on histogram distance functions. Results show that the proposed algorithm required a learning dataset size as small as 5 samples and was resistant to learning labelling error up to 50%. |
---|