| Abstract [eng] |
In this work, several different architectures of convolutional neural networks are being investigated, trying to adapt them to quality assurance of industrial production by recognizing surface defects. Based on the research work presented in the sources with an analogous database, a model is first created and the accuracy of the results, obtained by changing the dimensions of the input images, is investigated. The second part of the study attempts to adapt the models of convolutional networks that have already been learned on the ImageNet database. Finally, in the last part of this research thesis, the defect localization functionality is created and the U-Net architecture is adapted to the database being studied in order to obtain defect segmentation maps. The model comparison methodology includes several different criteria including the accuracy achieved, the complexity and speed of the training process, and ultimately the functionality for defect localization or segmentation. The results of the study showed that it was easier to train models into which ImageNet weights were imported than to create new models as suggested in the related works, but the proposed defect localization function and training speed were not good enough compared to the segmentation results achieved by U-Net. |
