Speaker
Description
Surface defects on galvanized strips are a key factor for the product quality in the flat-rolled steel industry. Some of the most notorious examples are slivers and scale defects. Correct classification and reduction of sliver and scale defects are essential for high end-product quality, lower downgrading rates, less CO2 emissions, and consequently lower costs.
Steelmakers currently use automated surface inspection systems (ASIS) from various vendors to inspect the products after hot rolling, pickling, and galvanizing. These typical inspection systems use conventional classifiers and as a result, their accuracy is limited. To circumvent this limitation the manufacturers are increasingly switching to defect classifiers that are based on neural networks.
Smart Steel Technologies (SST) uses multiple deep neural networks that are specialized for each process step. In addition to the surface inspection data, SST also utilizes the Level 1 and Level 2 production data to improve classification accuracy.
Neural networks are known for requiring a lot of hand-labeled images to perform well. In this paper, we focus on different aspects of effective data collection and emphasize the importance of the continuous involvement of domain experts in this process. We also highlight the effectiveness of the cross-process, multi-view approach to improve classification accuracies.
Keywords
defect image classification, surface inspection, deep learning, convolutional neural networks, cross-process classification, sliver defects, scale defects