Jointly Class-specific and Shared Discriminative Dictionary Learning for Classifying Surface Defects of Steel Sheet

抄録

For surface defect images that captured from a practical steel production line, feature extraction for classification is a challenging task due to inter-class similarity and intra-class difference of defect images. In this paper, we propose a class-specific and shared discriminative dictionary learning (CASDDL) method for extracting discriminative class-specific features to classify surface defects of steel sheet. Specifically, we learn a shared sub-dictionary as well as several class-specific sub-dictionaries to explicitly capture common information shared by all classes and class-specific information belonging to corresponding class. We adopt a mutual incoherence constrain for each sub-dictionary to encourage learned sub-dictionaries to be as independent as possible. A Fisher-like discriminative criterion is also introduced to coding vectors over all the class-specific sub-dictionaries, which can indirectly improve the discriminative ability of learned dictionary. In addition, we further impose a low-rank constrain on coding vector over shared sub-dictionary to guarantee learned dictionary has the abilities of same class of samples itself reconstruction and different class of samples reconstruction. Finally, classification can be efficiently performed by discriminative coding vector that obtained from a reconstructive and discriminative dictionary. Experimental results demonstrate that our proposed CASDDL method achieves a comparable or better performance than the state-of- the-art dictionary learning methods in classifying surface defects of steel sheet.