Neural Network Style Transfer of Defects from Concrete to Metal to Improve Monitoring Efficiency

The article is devoted to the study of dependence of the quality of defect detectors on metal products on the augmentation of the initial dataset. Special attention is paid to a non-standard method of augmentation - the method based on style transfer. For 2000 reference metal images, crack formation was generated from concrete images. The analysis showed that such an extension of the training sample leads to 2-3% accuracy improvement on the test sample. Various architectures have been explored. Converged dual-pass detectors, single-pass detectors and transformer detectors. All the architectures show an accuracy gain using the proposed augmentation.