Visual inspection of steel surface defects based on domain adaptation and adaptive convolutional neural network

Abstract Automatic inspection methods based on machine vision have been widely employed for steel surface defect detection. The central purpose of these methods is to extract features to represent different defects. However, current methods depend on machine learning that demands handcrafted features and overlooks the domain shift. In this paper, we propose a new method combining domain adaptation (DA) and adaptive convolutional neural network (ACNN), called DA-ACNN, to achieve steel surface defect detection. The convolutional neural network (CNN) is used as the backbone. To account for the lack of labels in a new domain, we introduce an additional domain classifier and a constraint on label probability distribution to achieve the cross-domain and cross-task recognition. The normal distribution and the quadratic function are used to optimize the loss to improve the network performance. Adaptive learning rates based on the loss and the weight, respectively, are proposed to minimize the losses of DA and classification. We conducted experiments on steel surface defect datasets to validate the effectiveness of DA-ACNN. Compared with the classical CNN and other approaches, the results demonstrate the superiority of the proposed method.