Abstract:

The production of seamless hot long metal products starts from an initial raw steel billet that follows progressive transformations through subsequent mandrel mills that assure final dimensions and mechanical properties. However, the roller cages in the push benches cansporadically produce marks and defects on the steel surface that are difficult to detect in hot steel conditions of more than 1000ºC. The challenges for a correct detection of defects are not trivial because of the high speed acquisition, and the demanding and hazardous environmental conditions, with the presence of dirt, oil vapor and water vapor. The inherent spectral radiance which maximum biases towards shorter wavelengths with the temperature increment makes the acquisition part non-trivial. Based on these premises, Tecnalia developed the SURFIN surface quality inspection system that performs real time detection and classification of external defects (e.g. roll marks, cracks, etc.) on the hot steel product manufacturing process of long metallic products such as bars, tubes, billets, slabs, beam blanks or structural profiles. The equipment is installed in the production line. It can detect defects at early stages in the production process, when the product is still incandescent (>1000ºC) allowing preventing the unnecessary addition of value to it. In this work, we present the performance improvements, in terms of speed, scalability and surface defect detection and classification performance of SURFIN system by replacing the commercial software-based detection and classification module from the first SURFIN version –supported by opaque handcrafted feature extraction and Support Vector Machines (SVM)– with an in-house made candidate window detection stage and a Convolutional Neural Network (CNN) performing the actual defect classification (CNN-SURFIN). The image database used included 3886 cropped images from long hot bars (2475 good and 1411 showing three defect categories). The classification-stage evaluation showed that CNNSURFIN architecture-based classifier outperformed (AUC=0.9970) the commercial SVMbased classifier (AUC=0.88) by a large margin. Additionally, two classical computer vision classification schemes based on texture features (LBP) feeding a random forest classifier (AUC=0.95) and a support vector machine classifier (AUC=0.92) were tested. In terms of speed the new architecture produces a noticeable impact in the overall performance of the SURFIN system, drastically cutting down the required computation time. It will serve as an example the case of a part image whose dimensions are 1024 pixels wide by 270.000 pixels high. In the previous SVM architecture an average of 46 seconds were needed in order to analyze the whole image, whereas the CNN-SURFIN classifier provides the outcome in about 21 seconds. The system is currently going through extensive validation and tests at steel production line 1 at Olarra SteelWorks factory, while new ongoing developments aim at an end-to-end trainable detection scheme.

Bibtex:

@inproceedings{picon_high_2017,
	address = {Bilbao, Spain},
	title = {High speed quality inspection of hot long metal products surface based on {Deep} {Convolutional} {Neural} {Networks}},
	booktitle = {Open session - {International} {Summer} {School} on   {Deep} {Learning} 2017},
	author = {Picon, Artzai and Alvarez-Gila, Aitor and Duro, Gorka and Cruz-Lopez, Antonio and Linares, Miguel and Lago, Alberto and Garrote, Estibaliz and Gutierrez-Olabarria, Jose A.},
	month = jul,
	year = {2017}
}