Multitask Learning for Unified Optical Surface Inspection: A Single Shared-Encoder Architecture for Thermal Reconstruction, Phase Unwrapping, and Defect Detection

Abstract

Current deep learning approaches to optical surface inspection deploy separate specialized models for each measurement task—individual networks for thermal image reconstruction, phase unwrapping, and defect detection. This modular architecture creates inefficiencies: each model requires its own training pipeline, computational resources, and maintenance burden, and the models cannot share the visual representations they learn about common optical surface features. This study proposes a multitask learning framework for optical surface inspection that employs a single unified architecture with a shared encoder and task-specific decoders to perform thermal reconstruction, phase unwrapping, and defect detection simultaneously. Built upon the deep learning 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 proposed unified model exploits the structural and physical commonalities across optical inspection tasks—the shared representation of surface geometry, material properties, and defect signatures—to achieve both computational efficiency and improved generalization through inter-task knowledge transfer. A comprehensive evaluation on synthetic and real optical measurement data demonstrates that the unified model achieves performance within 3.8% of task-specific specialist models while reducing total model size by 62% and inference compute by 58%. The shared encoder learns cross-task representations that improve each individual task's accuracy relative to single-task training, validating the presence of positive knowledge transfer across tasks. The framework provides a practical and theoretically principled pathway toward unified, efficient optical inspection systems.