Asymmetric Labeling for Defect Inspection Prediction Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing defect inspection systems using machine learning models often result in false negative recognition, where defective products are incorrectly identified as normal, due to the limitations of training data and classification methods.
Innovation Solution
The system employs a prediction model trained with asymmetric ground truth labels for normal and defective products, where normal products are labeled only as normal and defective products are labeled with multiple defect types and weights, to minimize the loss value and reduce false negative recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a learning model is trained using conventional symmetric ground truth labels (where normal and defective products are labeled with equal weight), then the model achieves balanced classification performance, but false negative recognition occurs where defective products are incorrectly identified as normal products
Solution Approach 1:
The patent applies asymmetry by introducing asymmetric ground truth labels where defective products are assigned higher weights than normal products. Specifically, defective products receive a weight of 1.0 while normal products receive a weight of 0.1, creating an asymmetric loss function that prioritizes correct identification of defective products over maintaining balanced classification performance
Solution Approach 2:
The patent changes the parameter weights assigned to different product types in the ground truth labels. By adjusting the weight parameter from equal values (symmetric) to differentiated values (asymmetric, with defective products having higher weights), the loss function becomes more sensitive to misclassification of defective products, thereby reducing false negatives
2Measurement precision
If a learning model is trained to minimize overall classification error, then general accuracy is improved, but the model fails to prioritize detection of defective products
Solution Approach 1:
The patent modifies the parameter weights in the ground truth labels to reflect the relative importance of different product types. Defective products are assigned a weight of 1.0 while normal products are assigned 0.1, which changes the optimization objective of the learning model to prioritize correct defect detection over overall accuracy
3Productivity
If conventional training data with uniform labeling is used, then the training process is simple and fast, but the model cannot distinguish the critical importance of defect detection
Solution Approach 1:
The patent introduces weighted parameters in the ground truth labels without changing the fundamental training process. The asymmetric weights (1.0 for defective, 0.1 for normal) are incorporated into the loss function calculation, allowing the model to learn defect prioritization while maintaining the efficiency of standard supervised learning procedures
Data Source
AI summary
As training data that is used in generation of a prediction model, training data for a normal product which is configured by assigning a normal product ground truth label including only a normal label indicating a possibility of correspondence to a normal product to a learning image of the normal product, and training data for a defective product which is configured by assigning a defective product ground truth label including only a plurality of weighted defect type labels indicating a possibility of correspondence to a plurality of defect types to a learning image of the defective product are used. According to this, it is possible to perform defect inspection with the prediction model in which a possibility of erroneously predicting the defective product as the normal product is further reduced by setting a loss value in a case of prediction as the normal product from a learning image of the defective product to which the defective product ground truth label is assigned to be larger than a loss value in a case of prediction as the defective product in a defect type other than a ground truth from the same learning image in machine learning.