SteelDefectClassicationwith.PDF

Steel Defect Classi?cation with Max-Pooling Convolutional Neural Networks Jonathan Masci, Ueli Meier, Dan Ciresan, Gabriel Fricout Jurgen¨ Schmidhuber Arcelor Mittal IDSIA, USI and SUPSI Maizieres` Research SA, Galleria 2, 6928 Manno-Lugano, France Switzerland {gabriel.fricout@} {jonathan, ueli, dan, juergen}@idsia.ch Abstract—We present a Max-Pooling Convolutional Neural si?cation. The system is usually based on a set of hand- Network approach for supervised steel defect classi?cation. On a wired pipelines with partial or no self-adjustable parameters classi?cation task with 7 defects, collected from a real production which makes the ?ne-tuning process of this industrial systems line, an error rate of 7% is obtained. Compared to SVM classi?ers trained on commonly used feature descriptors our best cumbersome, requiring much more human intervention than net performs at least two times better. Not only we do obtain desired. In this work we focus on the two last pipeline stages much better results, but the proposed method also works directly and propose an approach based on Max-Pooling Convolutional on raw pixel intensities of detected and segmented steel defects, Neural Networks (MPCNN) [1], [2], [3], [4], [5], that learn avoiding further time consuming and hard to optimize ad-hoc the features directly from labeled images using supervised preprocessing. learning. We show that the proposed method achieves state- of-the-art results on real world data and compare our approach I. INTRODUCTION to classi?ers trained on clas