Defected Object Detection Through One-Class Support Vector Machine Exploiting Deep Features.

In the context of Industry 4.0, a fundamental role is covered by the detection of defected or damaged components in industrial processing lines. A huge amount of methods and approaches have been designed during the last decades, in order to achieve better and better detecting performance. During last years, automated visual inspection systems based on image processing analysis and neural networks became really diffused and efficient and an impressive number of different approaches have been explored, particularly after that deep neural networks consolidated their extremely efficiencies into extract complex features from image data. In this field, one-class classification methods (i.e., methods that recur only to good components training data) are acquiring more and more interest due to the fact that in many cases very limited data is available for the defected products. In this work, a one-class classification approach based on Support Vector Machine which exploits deep neural networks for feature extraction is explored in relation to a real industrial production case, bringing the system to an embedded platform and showing so a Proof-of-Concept for a final on-field application.