Event Details
Addressing Training Data Limitations: A Case Study of Inspection in Automated Fiber Placement
Presenter: Assef Ghamisi
Supervisor:
Date: Mon, August 21, 2023
Time: 10:00:00 - 00:00:00
Place: ZOOM - Please see below.
ABSTRACT
Zoom Details:
Meeting link: https://uvic.zoom.us/j/84535210721?pwd=T2RRSzRiOThkM1d4eER0N1luNmwxdz09
Meeting ID: 845 3521 0721
Password: 272486
Note: Please log in to Zoom via SSO and your UVic Netlink ID
Abstract: This work presents visual defect detection methods that tackle data challenges such as labeling, scarcity, and imbalance. The proposed methods are demonstrated within an Automated Fiber Placement (AFP) case study, specifically utilizing depth map images of the composite surface. The first method employs unsupervised anomaly detection, effectively spotting defects using only non-defective samples. It involves boundary estimation for each composite strip (tow), followed by a sliding window that extracts image patches. These patches are processed through an autoencoder trained on normal samples, revealing anomalies via reconstruction errors. Aggregating the values of reconstruction error creates an anomaly map, and defects are pinpointed by performing blob detection on this map. This approach achieves strong accuracy in binary classification and defect localization. The second method relies on rule-based computer vision, without needing training data. It detects gaps and overlaps, major defects in AFP. The proposed framework identifies outlines of the composite tapes, enabling comparison between consecutive tows to identify any gaps or overlaps. Comparing detected defects to human