Synthetic Data Generation for Steel Defect Detection and Classification Using Deep Learning

Aleksei Boikov,Roman Savelev,Alexandr Kolesnikov,Vladimir Payor

doi:10.3390/sym13071176

Abstract

The paper presents a methodology for training neural networks for vision tasks on synthesized data on the example of steel defect recognition in automated production control systems. The article describes the process of dataset procedural generation of steel slab defects with a symmetrical distribution. The results of training two neural networks Unet and Xception on a generated data grid and testing them on real data are presented. The performance of these neural networks was assessed using real data from the Severstal: Steel Defect Detection set. In both cases, the neural networks showed good results in the classification and segmentation of surface defects of steel workpieces in the image. Dice score on synthetic data reaches 0.62, and accuracy—0.81.

Highlights

Machine learning algorithms for computer vision are widely used in various industries
This study considers the possibility of applying synthesized data for semantic segmentation and classification of defects in steel products
There are the results of generating synthetic data for different types of defects

Summary

Introduction

Machine learning algorithms for computer vision are widely used in various industries. The quality of a machine-learning-based automation system largely depends on the quality of the initial training sample. It should maximally reliably reflect the nature of the process under study, in other words, be representative [1]. Obtaining such a sample is very laborious; it is necessary to capture as many different variants of the object states under investigation as possible [2]. This may cause difficulties because of the intraclass variation of the object, i.e., objects belonging to the same class may have a different representation (color, shape, size, etc.) [3]

Methods

Results

Discussion

Conclusion