HIGH-THROUGHPUT DEEP LEARNING ALGORITHM FOR DIAGNOSIS AND DEFECTS CLASSIFICATION OF WATERPROOFING MEMBRANES

Abstract

The work is devoted to the development of a high-performance deep learning algorithm related to the diagnosis and classification of defects of water-repellent membranes. The mechanism of constructing visual models of the membrane surface is discussed. This allows to get the representative training data set. The proposed methodology consists in the sequent transformations of pixel-image intensities to find defected fragments on the membrane's surface. The computational algorithm is based on the architecture of convolution neural networks. To assess its effectiveness, the "confidence of confidence" criterion is proposed. The presented computations show that the methodology can be successfully applied in material sciences, for example, to study the properties of building materials, or in forensic science when examining the causes of construction catastrophes.