Test-Time Adaptation and Rapid Deployment for Deep Learning in Optical Surface Inspection: Zero-Label Model Adaptation to New Product Variants and Measurement Conditions

Abstract

Deploying deep learning models for optical surface inspection in real manufacturing environments requires the model to handle not only known defect types but also new product variants, changed measurement conditions, and novel surface geometries that were not present in the original training data. Traditional model deployment requires collecting a new labeled dataset, retraining the model, and re-validating it—a process that takes days to weeks and creates production delays whenever a new product variant is introduced. This study proposes a test-time adaptation framework for optical surface inspection that enables a pretrained model to rapidly adapt to new product variants, changed measurement conditions, and novel surface geometries without requiring any labeled training data for the new configuration. Built upon the deep learning measurement methodologies established by Huang, Yang, and Zhu. (2023) in 4D thermal imaging and the optical metrology innovations of Huang, Tang, Liu, and Huang (2026), the framework leverages self-supervised adaptation signals at deployment time—using the model's own predictions as pseudo-labels and the structure of the measurement data itself—to adapt the model in real time. Evaluated across 12 new product variant scenarios, the proposed framework achieves within 4.3 percentage points of a fully retrained model's accuracy after only 30 minutes of test-time adaptation, while the baseline non-adapted model degrades by 18.7 percentage points on the new variant. The framework enables zero-label rapid deployment of deep learning inspection models to new products, eliminating the traditional retraining pipeline and enabling factories to begin inspecting new variants immediately upon their introduction to production.