Neural models application for image recognition in the field of industrial automation

Artificial Intelligence is emerging as a disruptive technological field, with the potential to revolutionize the industrial world. In this context, Machine Learning is generally referred to as a branch of techniques which are designed to allow machines to learn and extract patterns from data. Among these, the so-called Deep Learning models have received particular interest for their potential to reduce human operations in several different applications. In this thesis, the attention is focused on the development and applicability of these models to image recognition and classification for industrial automation. In particular, the considered test case focuses on the automation of the quality control of brake calipers, from an important automotive player, in a typical production line of a company specialized in industrial automation. Particularly, Neural Network models are developed and trained, on a proper database, to be able to screen and identify different defects which may be present (such as imperfections, scratches or bubbles on the painted surface). All defects were artificially reproduced on brake callipers by means of CAD renderings, whose generation was automated via Python scripts to generate large enough databases for training of the Neural models. This artificially-generated database is particularly important for the application of the developed algorithms, as it remarkably reduce the time and effort to acquire real images of the calipers. The Neural Networks were then built using Python code in the Google Colab environment and trained based on the generated database. In order to improve the model accuracy and reduce overfitting, different models have been tested to identify that with suitable characteristics to guarantee the best performance for the considered case. The developed methodology is envisioned to be relevant for application to the industrial automation of the considered test case, reducing manual operations as well as for improving the energy efficiency of the production line. This perspective is particularly interesting in the context of the industrial digitalization and Industry 4.0 goals.