Partnerzy

Streszczenie:
Steel surface defects in both flat and long products are undesired not only from an aesthetic point of view, but also can lead to severe deterioration of material performance. Manual defect inspection is slow and costly, and thus, automatization of such processes is of interest. Several steel surface defect datasets have been made publicly available so far, and the most famous of them is the Northeastern University (NEU) surface defect database. Many research on surface defect inspection has already been conducted using this dataset, and excellent prediction capabilities were demonstrated in the open literature. More recently, this dataset was extended to account for effects that are expected to occur in real industrial scenarios, such as motion blur, non-uniform illumination, and noise. The extended dataset containing images with those modifications was also made publicly available (E-NEU). In previous papers on the subject, it was shown that using deep learning models trained on the NEU dataset to the E-NEU dataset does not necessarily lead to correct predictions. In this paper, based on the steel surface defects analysis, it is demonstrated that the performance of deep learning architectures can be effectively improved by applying image preprocessing techniques

Słowa kluczowe:
Surface defects classification, Quality control, Steel Surface, Long products, Flat products