A Novel Methodology for Vision Backbone Network Fine-Tuning and Continual Learning in Optical Inspection Tasks

Abstract

Roll-to-roll (R2R) manufacturing plays a key role in the
production of electronic devices, liquid-crystal films, and
other advanced materials. In these processes,
maintaining precise material thickness and surface
uniformity is essential to ensure product quality, process
efficiency, and yield. While optical inspection systems are
widely used to detect surface defects, using AI to
automatically segment and classify the defect remains a
challenging task. This paper investigates a new
methodology to establish an AI powered optical
inspection testbed that automatically handles multi-label
defect classification and semantic segmentation in R2R
manufacturing environments. The proposed system
integrates semantic segmentation and multi-label defect
classification into an automated optical inspection
pipeline. Using a Vision Transformer deep learning
architecture, our classification model achieved a high
accuracy across multiple surface defect categories in a
production-representative dataset. Our framework
consists of automatic image pre-processing, rapid fine
tuning of pre-train vision models, and a continuous
learning strategy after the model is deployed. The system
is intended to handle high-throughput image data with
minimal latency, enabling fast and accurate product
inspection while maintaining a light-weight model
architecture. The end-to-end framework is designed to
enhance in-line defect detection, reduce human
dependency, and improve overall process visibility. The
proposed technology has been successfully deployed to a
manufacturing plan and has achieved satisfactory
detection and classification performance in a real
operation environment.