Classification of Defects in Welds Using a Convolution Neural Network

In this paper, the possibility of using a convolutional neural network for analyzing weld defects are considered. A program has been developed for creating a data set for training and classification of weld defects based on the publicly available GDXray database. The VGG-16 model is used for classification. The model and algorithm for training a convolutional neural network are described. Technologies of model retraining control are applied. The main method for solving the problem is to transfer training due to the small number of images for training. The resulting model was applied in a specially created program for detecting and classifying weld defects. Examples of how the program works are presented. This model classifies weld defects into 5 classes with an average accuracy of about 86%.