Vision inspection with neural networks

This work describes a system for extracting and classifying defects inside bottles for cosmetic and pharmaceutical use. The system integrates various defects identification and automatic classification algorithms based on neural networks (NN). The aim is to be able to identify defective bottles at the end of production chain. In a set of 60 bottles, 3600 images were taken, 60 for each. We extracted 4161 defects of which, 70% was used for training and 30% for testing the neural network. We considered five defect classes (rubber, aluminum, glass, hair and tissue) with more than 90% accuracy on the test set.