ArcDef is Loss Function for Cracks Classification

This paper proposes the use of a new loss function to train a defect classifier on images of metal structures. The developed function allows to vary different classes in the hyperplane space in such a way as to provide the highest classification accuracy. An advantage of 1.3% compared to known models on the dataset of PJSC “Severstal” is shown. A comparative analysis for different neural network architectures was performed. The best results were shown by Vision Transformer, which achieved 94.9% on the F1-score metric.