Asymmetric Loss Function for Defect Detection Training Data Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current supervised learning algorithms for defect detection in manufacturing face challenges due to erroneous human labeling and high turnaround time, leading to mispredictions and difficulty in diagnosing labeling errors, as they penalize predictions in defect-free regions and fail to accurately model unlabeled defects.
Innovation Solution
A method and system that refine the training pipeline by using a processor to receive and process training images, creating a first model with a specific loss function that penalizes predictions in negative images but not in positive images, allowing for human review and iteration to create a second model for improved object detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If supervised learning algorithms penalize predictions in defect-free regions to avoid false positives, then manufacturing precision improves, but measurement precision deteriorates because the algorithm cannot distinguish between actual defects and labeling errors
Solution Approach 1:
The loss function applies different quality standards to different regions: defective regions allow additional predictions without penalty while defect-free regions maintain strict penalization. This local differentiation enables the model to tolerate labeling errors in defective regions while maintaining high precision in defect-free regions.
Solution Approach 2:
Instead of uniformly penalizing all additional predictions, the patent inverts the traditional approach by selectively not penalizing additional predictions in defective regions. This inversion allows the model to recover from labeling errors while maintaining precision in clean regions.
2Manufacturing precision
If human reviewers manually label all defects in training images to improve data quality, then manufacturing precision improves, but productivity deteriorates due to high time consumption
Solution Approach 1:
The system enables partial self-service by allowing the model to identify additional defects in positive images without human intervention. The asymmetric loss function automatically handles labeling errors by not penalizing additional predictions in defective regions, reducing the need for manual review while maintaining data quality.
Solution Approach 2:
Instead of requiring complete manual labeling of all defects, the patent applies partial action by only requiring human review when necessary. The asymmetric loss function handles the majority of labeling errors automatically, allowing the system to achieve high data quality with reduced human effort.
3Device complexity
If supervised learning algorithms use traditional loss functions that penalize all additional predictions, then device complexity remains low, but reliability deteriorates due to mispredictions from unlabeled defects
Solution Approach 1:
The asymmetric loss function introduces local quality by applying different penalty rules to different image types. This localized approach maintains relative model simplicity while significantly improving reliability by preventing mispredictions caused by labeling errors in defective images.
Data Source
AI summary
Embodiments of the present disclosure discuss a system [100] and a method [200] for refining training images for object detection. Conventional systems provide inaccurate predictions of defects because of erroneous labeling of defects in training images. Further, existing systems focus on improving the algorithms used for object detection. Embodiments of the present disclosure address these problems by using supervised learning techniques and refining the training data provided to train the algorithms. The system [100] refines the training images using predictions of a first model and providing the predictions for human review, and subsequently uses the refined training images to create a second model that can be used in a production environment of an enterprise for object detection.