Využití metod strojového učení pro testování kvality průmyslových výrobků

Abstract

The main task of master thesis is the analysis of machine learning for visual inspection of products. Specifically, evaluation of surface defects on products. The field of machine learning includes a wide range of algorithms, but for the master thesis it was decided to focus only on the field of deep learning and convolutional neural networks. At the beginning, a search was made of the area of surface defect inspection using neural networks. This was followed by an analysis of the tools used for the development of neural networks and related libraries. The last part was a practical implementation in the form of creating a model for a segmentation task and training on a dataset of SMD components. In this work, two approaches are compared. The first was the use of the commercial software Cognex Deep Learning Studio and its tools for training deep learning models. And the second approach was to implement U-Net and SegDecNet architectures using the TensorFlow library. In the case of the U-Net architecture, an additional test was proposed, consisting in comparing the results of training on whole images or on images that were divided into smaller parts with a defined overlap. The Cognex Deep Learning Studio was used to annotate the SMD component dataset. The result is evaluation metrics that compare the model from the commercial Cognex Deep Learning Studio and the implemented U-Net and SegDecNet architectures.