Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adaptive few-shot learning for robust part quality classification in two-photon lithography

Sixian Jia, Ruo-Syuan Mei, Chenhui Shao

TL;DR

The paper tackles the challenge of dynamic quality control in two-photon lithography by developing an adaptive CV framework that combines novelty detection, few-shot incremental learning, and few-shot domain adaptation on a scale-robust backbone. It introduces a 1D LDA-based novelty test with batch voting, a rehearsal-based two-stage incremental learning process, and a few-shot DANN for domain transfer between hemisphere and cube geometries, achieving 99–100% novelty-detection accuracy, ~92% incremental-learning accuracy with 20 shots, and 96.19% cube-domain accuracy with 5 shots. The approach demonstrates strong data efficiency and robustness to domain gaps, enabling deployment and maintenance of quality models in evolving TPL manufacturing environments. These results indicate significant practical impact for inline quality control and rapid adaptation to new defect classes and part designs in micro-scale AM processes.

Abstract

Two-photon lithography (TPL) is an advanced additive manufacturing (AM) technique for fabricating high-precision micro-structures. While computer vision (CV) is proofed for automated quality control, existing models are often static, rendering them ineffective in dynamic manufacturing environments. These models typically cannot detect new, unseen defect classes, be efficiently updated from scarce data, or adapt to new part geometries. To address this gap, this paper presents an adaptive CV framework for the entire life-cycle of quality model maintenance. The proposed framework is built upon a same, scale-robust backbone model and integrates three key methodologies: (1) a statistical hypothesis testing framework based on Linear Discriminant Analysis (LDA) for novelty detection, (2) a two-stage, rehearsal-based strategy for few-shot incremental learning, and (3) a few-shot Domain-Adversarial Neural Network (DANN) for few-shot domain adaptation. The framework was evaluated on a TPL dataset featuring hemisphere as source domain and cube as target domain structures, with each domain categorized into good, minor damaged, and damaged quality classes. The hypothesis testing method successfully identified new class batches with 99-100% accuracy. The incremental learning method integrated a new class to 92% accuracy using only K=20 samples. The domain adaptation model bridged the severe domain gap, achieving 96.19% accuracy on the target domain using only K=5 shots. These results demonstrate a robust and data-efficient solution for deploying and maintaining CV models in evolving production scenarios.

Adaptive few-shot learning for robust part quality classification in two-photon lithography

TL;DR

The paper tackles the challenge of dynamic quality control in two-photon lithography by developing an adaptive CV framework that combines novelty detection, few-shot incremental learning, and few-shot domain adaptation on a scale-robust backbone. It introduces a 1D LDA-based novelty test with batch voting, a rehearsal-based two-stage incremental learning process, and a few-shot DANN for domain transfer between hemisphere and cube geometries, achieving 99–100% novelty-detection accuracy, ~92% incremental-learning accuracy with 20 shots, and 96.19% cube-domain accuracy with 5 shots. The approach demonstrates strong data efficiency and robustness to domain gaps, enabling deployment and maintenance of quality models in evolving TPL manufacturing environments. These results indicate significant practical impact for inline quality control and rapid adaptation to new defect classes and part designs in micro-scale AM processes.

Abstract

Two-photon lithography (TPL) is an advanced additive manufacturing (AM) technique for fabricating high-precision micro-structures. While computer vision (CV) is proofed for automated quality control, existing models are often static, rendering them ineffective in dynamic manufacturing environments. These models typically cannot detect new, unseen defect classes, be efficiently updated from scarce data, or adapt to new part geometries. To address this gap, this paper presents an adaptive CV framework for the entire life-cycle of quality model maintenance. The proposed framework is built upon a same, scale-robust backbone model and integrates three key methodologies: (1) a statistical hypothesis testing framework based on Linear Discriminant Analysis (LDA) for novelty detection, (2) a two-stage, rehearsal-based strategy for few-shot incremental learning, and (3) a few-shot Domain-Adversarial Neural Network (DANN) for few-shot domain adaptation. The framework was evaluated on a TPL dataset featuring hemisphere as source domain and cube as target domain structures, with each domain categorized into good, minor damaged, and damaged quality classes. The hypothesis testing method successfully identified new class batches with 99-100% accuracy. The incremental learning method integrated a new class to 92% accuracy using only K=20 samples. The domain adaptation model bridged the severe domain gap, achieving 96.19% accuracy on the target domain using only K=5 shots. These results demonstrate a robust and data-efficient solution for deploying and maintaining CV models in evolving production scenarios.
Paper Structure (29 sections, 6 equations, 19 figures, 3 tables)

This paper contains 29 sections, 6 equations, 19 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Backbone Model
  4. Hypothesis Testing
  5. Dimensionality Reduction via LDA
  6. Novelty Scoring
  7. Sample Thresholding and Vote Calibration
  8. Few-Shot Learning for Incremental Learning
  9. Incremental Update Strategy
  10. Two-Stage Fine-Tuning
  11. Few-Shot Learning for Domain Adaption
  12. Dataset
  13. Results and Discussion
  14. Hypothesis Testing
  15. LDA Feature Space Separation
  16. ...and 14 more sections

Figures (19)

  • Figure 1: Overview of the proposed backbone model.
  • Figure 2: The multi-stage hypothesis testing pipeline. Features from the backbone are projected to 1D using LDA, then scored. A batch-level vote based on these scores determines if the batch belongs to a new class.
  • Figure 3: Architecture of the proposed DANN.
  • Figure 4: Sample images from the two fabrication domains. Hemisphere domain: (a1) good; (a2) minor damaged; and (a3) damaged. Cube domain: (b1) good; (b2) minor damaged; and (b3) damaged.
  • Figure 5: 1D LDA projections with minor damaged as the new class. (a) validation with known good samples; (b) novelty detection of unknown minor damaged samples; (c) validation with known damaged samples.
  • ...and 14 more figures