Few-Shot Visual Inspection Meta-Learning for Cross-Domain Defects
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning-based computer vision algorithms for visual inspection require significant data and time to adapt to new domains, and methods like transfer learning and domain adaptation do not generalize well with limited data, especially across diverse domains.
Innovation Solution
A meta-learning system using squeeze-excitation modules, anti-aliasing filters, classification loss, and contrastive loss, along with snapshot ensembling and Self-Optimal Transport (SOT) feature transform, enables rapid adaptation to new visual inspection tasks with minimal labeled examples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional deep learning methods are used for visual inspection, then detection accuracy can be achieved, but significant amounts of labeled data and training time are required for each new domain
Solution Approach 1:
The system performs preliminary action by pre-training the deep learning model on multiple source domains before deployment. This meta-training phase prepares the model to quickly adapt to new target domains with minimal data, resolving the contradiction by reducing the labeled data requirement while maintaining detection accuracy through prior knowledge acquisition
2Measurement precision
If traditional deep learning methods are used for visual inspection, then detection accuracy can be achieved, but significant training time and computational resources are required for each new domain
Solution Approach 1:
The system performs preliminary action by pre-training the deep learning model on multiple source domains before deployment. This meta-training phase prepares the model to quickly adapt to new target domains with minimal data, resolving the contradiction by reducing the labeled data requirement while maintaining detection accuracy through prior knowledge acquisition
Solution Approach 2:
The system applies parameter changes by modifying the training regime from domain-specific training to meta-learning across multiple domains. By changing the optimization objective to learn domain-invariant features and using techniques like domain adaptation, the model achieves fast adaptation to new domains with reduced training time while maintaining detection accuracy
3Productivity
If transfer learning or domain adaptation methods are used, then adaptation to new domains is faster, but generalization to diverse domains with limited data is poor
Solution Approach 1:
The system applies universality by designing a domain-agnostic visual inspection framework that can handle multiple diverse domains (semiconductors, displays, solar panels, etc.) with a single model architecture. The model learns universal defect detection capabilities across domains while maintaining the ability to adapt to domain-specific characteristics, thus achieving both fast adaptation and broad generalization
Data Source
AI summary
Systems and methods described herein which can involve for a first input of a plurality of labeled images of a new domain task, processing the first plurality of labeled images through a plurality of backbone snapshots, each of the backbone snapshots representative of a model trained across a plurality of other domain tasks, each of the plurality of backbone snapshots configured to output a first plurality of features responsive to the input; processing a second input of second plurality of unlabeled images through the plurality of backbone snapshots to output a second plurality of features responsive to the second input; and generating a representative model for the new domain task from the clustering and transformation of the first plurality of features and as associated from the clustered and transformed second plurality of features.