Machine Learning methods for image analysis and their application to industrial optimization problems

Machine Learning (ML) techniques are becoming every day more effective for the analysis of large datasets in a wide range of engineering applications. This thesis particularly focuses on ML techinques for image analysis, that is, those techniques which allow data extraction from images. In order to understand the foundations of the subject of study, a general overview of the available techniques for the target application is first presented, along with a discussion of the related theoretical aspects. Second, a practical application of these techniques is targeted: the analysis of defect recognition in an industrial production process. To this, a complete workflow which includes image pre-processing, training of the ML models, and post-processing of the results, is developed. The workflow has been extensively tested on a real industrial test case, and the results show that proper tuning of the ML training allows to obtain accurate results in terms of defect recognition on real manufactured parts. All the analysis related to the tuning of the workflow and to the results obtained with different approaches are reported and discussed in detail. The final version of the developed workflow is shown to properly perform for the target problem, thus, in perspective, it could be implemented for an ML powered version of the considered industrial problem. This would finally allow to achieve better performance of the manufacturing process and, ultimately, to achieve a better management of the available resources in terms of time consumption and energy expense.