Knowledge Distillation for Optical Surface Inspection: Compressing Large Inspection Networks into Efficient Deployable Models

Abstract

State-of-the-art deep learning models for optical surface inspection achieve impressive accuracy but are computationally expensive, requiring high-end GPU servers for real-time inference. This creates a deployment barrier for manufacturing facilities that lack GPU infrastructure or require inspection systems that can operate on low-power embedded devices. Knowledge distillation addresses this problem by training a compact student network to mimic the behavior of a larger teacher network, transferring not just the final predictions but the rich dark knowledge encoded in the teacher's logits, attention maps, and intermediate representations. This study proposes a comprehensive knowledge distillation framework for optical surface inspection that compresses high-accuracy teacher models into lightweight student networks suitable for edge deployment, while preserving the teacher's ability to detect rare defects, reason about uncertainty, and generalize across diverse product variants. 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 combines logit-based distillation, intermediate representation matching, and defect-aware prioritization to achieve up to 12.7× model compression while retaining 96.4% of the teacher's accuracy on defect detection and within 0.3 K of the teacher's thermal reconstruction MAE. The distilled student models achieve real-time inference at 156 FPS on a mobile ARM processor (Jetson Nano) and 94 FPS on a low-power edge TPU, enabling deployment of state-of-the-art inspection accuracy on compact, low-cost hardware. This work provides a practical pathway for deploying the most accurate optical inspection models on the full range of manufacturing hardware, from high-end datacenters to embedded edge devices.