SUALAB provides deep-learning inspection solutions for machine vision. It began with defect detection in textile printing, introducing machine vision to a process that lacked automated inspection in what the article describes as a world first.
SUALAB was acquired by Nasdaq-listed Cognex in a transaction described in the Korean article as approximately KRW 240 billion. This article examines one of its patents as an example of technology developed by a successful startup.
The patent described in this post is related to how to train an artificial neural network model.
[Patent]
Click the image to view the granted patent.
The patent proposes The patent uses a Weibull distribution to model feature distributions associated with normal and defective industrial products.Under the patent,
According to the patent, training encourages hidden-layer feature distributions to fit the specified Weibull distribution.This is incorporated into the weight updates.
Specifically, Where a hidden node's feature distribution resembles the target Weibull distribution, training encourages that node's activation because it may represent features useful for detecting defects.
Meanwhile, Where the distribution differs substantially from the target, training suppresses the node because it may be less useful for identifying defects.
This encourages the retained features to follow the target statistical patternand supports anomaly detection by the network.
This article reflects the information available when it was published. Contact us to discuss your circumstances.