Substrate Inspection Unit for Verified Deep-Learning Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In the deep learning process for substrate inspection during printing, unclear classification criteria and incorrectly labeled image data sets lead to reduced learning degree and prediction performance, particularly in the production of display devices like LCD and LED panels.
Innovation Solution
A substrate inspecting unit with modules for feature extraction, validity evaluation, class verification, and data reconstruction using pre-trained models and unsupervised learning techniques like t-SNE and DBSCAN to improve classification accuracy and reduce labeling work time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If manual labeling of image data is performed without verification, then labeling work time is reduced, but classification accuracy and prediction performance deteriorate
Solution Approach 1:
The system performs preliminary verification of training image data using unsupervised learning algorithms (t-SNE for dimensionality reduction, DBSCAN for clustering) before the deep learning training process. This preliminary action identifies and corrects mislabeled data, ensuring high classification accuracy while maintaining efficient labeling workflows.
Solution Approach 2:
The system implements a feedback mechanism where the deep learning model's prediction results are used to verify and refine the training data labels. The model processes training data, and the results feed back into the labeling verification process, allowing automatic correction of mislabeled images and continuous improvement of classification accuracy.
2Productivity
If deep learning training is performed with unverified image data, then training speed is improved, but prediction performance deteriorates
Solution Approach 1:
The system performs preliminary verification of training image data using unsupervised learning algorithms (t-SNE for dimensionality reduction, DBSCAN for clustering) before the deep learning training process. This preliminary action identifies and corrects mislabeled data, ensuring high classification accuracy while maintaining efficient labeling workflows.
Solution Approach 2:
The system implements a feedback mechanism where the deep learning model's prediction results are used to verify and refine the training data labels. The model processes training data, and the results feed back into the labeling verification process, allowing automatic correction of mislabeled images and continuous improvement of classification accuracy.
3Device complexity
If classification criteria are not clearly defined, then data processing simplicity is improved, but learning degree and verification reliability drop
Solution Approach 1:
The system employs unsupervised learning algorithms (t-SNE and DBSCAN) that automatically define classification criteria without requiring manual intervention. The algorithms self-organize the training data into clusters based on inherent patterns, automatically establishing clear classification boundaries and improving verification reliability while maintaining processing simplicity.
Data Source
AI summary
Provided are a substrate inspecting unit capable of reducing image data labeling work time through training image data set verification and semi-automatic image labeling, and at the same time, improving prediction performance by improving classification accuracy for data sets, and a substrate treating apparatus including the same. The substrate inspecting unit comprises a feature extracting module for extracting a feature from training data included in each class in response to a plurality of training data related to image data of a substrate being classified according to a predefined class, a validity evaluating module for evaluating validity of the feature, a class verifying module for verifying the predefined class, and a data reconstructing module for reconstructing the plurality of training data based on a feature determined as valid and a verified class, wherein reconstructed training data is utilized when inspecting the substrate.