Abstract:

Automatic inspection methods based on machine vision have been widely employed forsteel surface defect detection. The central purpose of these methods is to extract featuresto represent different defects. However, current methods depend on machine learning thatdemands handcrafted features and overlooks the domain shift. In this paper, we propose anew method combining domain adaptation (DA) and adaptive convolutional neural net-work (ACNN), called DA-ACNN, to achieve steel surface defect detection. The convolutionalneural network (CNN) is used as the backbone. To account for the lack of labels in a newdomain, we introduce an additional domain classifier and a constraint on label probabilitydistribution to achieve the cross-domain and cross-task recognition. The normal distribu-tion and the quadratic function are used to optimize the loss to improve the network per-formance. Adaptive learning rates based on the loss and the weight, respectively, areproposed to minimize the losses of DA and classification. We conducted experiments onsteel surface defect datasets to validate the effectiveness of DA-ACNN. Compared withthe classical CNN and other approaches, the results demonstrate the superiority of the pro-posed method.

This article is available online here and to CALCE Consortium Members for personal review.