| Abstract [eng] |
In the scientific work defectoscopy methods are analyzed, the image analysis method is analyzed in more detail. Deep neural networks are used as a tool to achieve the goal. The aim of this research is to investigate the possibility of using deep neural networks to detect surface defects in civil infrastructure from images. The methodological part describes the benefits of defectoscopy for industry and infrastructure, the evolution of image analysis, an overview of deep neural and U-Net networks. The evaluation estimates of network models and the method of transfer learning, database, its augmentation and the benefits of this method were then reviewed to achieve the ultimate goal. Artificial neural networks were trained in such a way that in the first case – when the image is submitted, the network recognizes the texture, in the second case – the network issues an answer whether the image has crack or not. In the third case – a U-Net type network is trained so that after submitting a surface image to the network input, the network finds and displays the defect. |
