Advanced Color Pseudo Anomaly to Enhance Learning of Convolutional Neural Network Models
Rizwan Ali Shah, HyungWon Kim · 2023
Advanced Vision Inspection Systems (VIS) enable automation of anomaly detection in the production lines of smart factories. Incorporating a pseudo-anomaly insertion step in the training process allows a target VIS to effectively differentiate normal and abnormal examples for specific product. However, prior research has not extensively explored the fundamental nature of abnormalities in the process of product defect detection. In this work, we introduce a novel Color Anomaly Insertion (CAI) scheme that adds anomalies into specific images considering color of anomaly as a key attribute. CAI transforms the color of anomalies into dark/bright and challenging forms. Overall anomaly insertion in CAI is done in two steps: first, anomaly patterns are extracted, and then their colors are transformed into desired color by altering RGB channels.