Surface defect identification of cross scene strip based on unsupervised domain adaptation

In view of the poor generalization performance of the deep learning model at surface defect identification of cross scene strip, an end-to-end multi-level aligned domain adaptation neural network (MADA) was proposed, which could achieve pixel-level illumination distribution alignment and feature-level texture distribution alignment, respectively. The source and target domain data were projected into the illumination subspace by MADA to achieve the pixel-level illumination distribution alignment, through the non-reference pixel-level illumination distribution alignment module and the illumination loss function. The adversarial learning of texture feature extractor and feature-level domain discriminator were used to achieve the texture distribution alignment of the source and target domain. The experiment achieved an F1 measure of 98% in Handan strip surface defect dataset and 86.6% in Severstal strip surface defect dataset. Experimental results showed that the proposed method has better generalization performance than other domain adaptation methods. © 2023 Zhejiang University. All rights reserved.