Optical inspection and standard computer vision techniques are today the two main ways of performing quality control in manufactoring. However, this two approaches limit the full automation of the production and impose solutions dependend on the specific products which translates to a higher cost with respect to a fully automatic approach. Moreover, standard machine vision algorithms always rely on ad-hoc parameters that have to be adjusted case by case with a loss of solidity and robustness.
In this thesis it was studied and tested an innovative approach to quality control on industrial products based on deep neural networks. A deep learning model does not requires parameters to be fixed in advance, it learns them from the data and uses them to classify a product as conforming or non-conforming. Moreover, coupling the model with a proper in-line acquisition system and a suitable harware, it was obtained that the monitoring of a producion line can be completely automatize.
The project has been carried out in collaboration with beanTech, IT company in Udine, and it is divided into two main parts: in the first one the attention is focussed on the implementation and testing of the deep learning model. The second one is more devoted to explore how to make the model ready to be integrated in a real production line, testing its performances, optmizing its workflow and making it compatible across multiple platforms.
In this kind of data-driven applications, great importance is attached to data preparation and pre-processing techniques as well as to classification accuracy and performance. A deep neural network can have hundreds of thousands or even milions of parametes and requires a large computational cost to be trained and executed.
Moreover, some manufactoring lines can require a rate of monitoring of milliseconds for each product. For these reasons, the use of dedicated and highly optmized libraries is mandatory, but also some low-level and custom optimizations can really have an impact on the performance of the application and, as it will be shown, sometimes they become crucial to make the model capable of achieving the task and being implemented in production.
Thesis not available.