Convolutional neural networks make Tecnalia’s SURFIN Hot surface inspection® system evolve to assure automatic quality control
Tecnalia has presented the new version of its SURFIN Hot surface inspection® system at the First European Machine Vision Forum of the European Machine Vision Association (EMVA). SURFIN performs in-line real-time detection and classification of surface defects (e.g. roll marks, cracks, etc.) from the manufacturing process of metallic products such as bars, tubes, billets, slabs, beam blanks or structural profiles. The systems are installed in the production line. It can detect defects at the early stages in the production process, when the product is still incandescent (>1000ºC). This allows preventing the unnecessary addition of value to it and having traceability of all the production, allowing a preemptive maintenance due to the information it obtains. The system is based on special 2D imaging with laser or LED-based imaging, and makes use of machine learning techniques.
SURFIN has been upgraded by replacing the previous detection and classification module –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). CNN-SURFIN architecture has been tested as a classifier over a database including 3886 cropped images from long hot bars (2475 good and 1411 showing three defect categories) in a 10-fold cross validation setup for good vs. bad classification, finding that it significantly outperformed (Area Under the ROC Curve – AUC=0.9970) the previous SVM-based classifier (AUC=0.88). We have also implemented two baselines by extracting texture features (LBP-Local Binary Patterns) and training an SVM (AUC=0.92) and a Random Forest classifier (AUC=0.95) on top of these, further supporting the superiority of the deep learning-based approach. Finally, CNN-SURFIN was trained for the full 4-class classification problem, yielding an AUC of 0.9956.
Having in mind that deep learning does not have the risk of model bias (we can easily accommodate an increasingly large number of training samples with high variance by adequately parametrizing it), and the bigger stability shown by the model, which makes it possible the training sets to be fed by a non-expert, this deep learning architecture makes the new SURFIN extremely flexible and accurate on complex hot steel classification tasks, and will be applied to several installations of the SURFIN system in operation.
A. Alvarez-Gila, A. Lopez-Cruz, S. Rodriguez-Vaamonde, M. Linares, J. A. Gutierrez-Olabarria, and E. Garrote, “Deep Convolutional Neural Networks for surface quality inspection of hot long metal products,” presented at the First European Machine Vision Forum, Heidelberg, Germany, 2016