Compact and Efficient Feature Representation for Defect in hot rolled strip steel: A postion paper.

This paper focuses on the compact and efficient feature representation for defect in hot rolled strip steel. Facing essential challenges of oil stain, water drops, steel textures and erratic illumination in the production line, we make a detailed analysis from two aspects of preprocessing and feature extraction. The Preprocessing improves image quality from the source, which includes homogeneous illumination compensation and pseudo-defect removal. Feature extraction is divided into traditional and deep-learning level, the advantages and disadvantages of the two sub-methods are compared. In addition, an original raw database CSU _STEEL is also introduced. Finally, the future research directions of feature representation for defect in hot rolled strip steel are prospected.