Continual learning for surface defect segmentation by subnetwork creation and selection

Automatic defects inspection plays an important role in product quality evaluation (Prunella et al., 2023). In the beginning of the field, the creation of meaningful features to find defective regions was done manually (Ojala et al., 2002; Chao and Tsai, 2008; Song and Yan, 2013; Jeon et al., 2014). Although classical machine learning methods have been proposed to identify images with defective surfaces (Jia et al., 2004; Agarwal et al., 2011; Shanmugamani et al., 2015), recent advances in deep learning research have led to an increase in performance (Prunella et al., 2023). Typically, there are three types of tasks for defect inspection with neural networks - classification, detection (He et al., 2019) and segmentation (Tabernik et al., 2020). In the case of defect classification, transfer learning helps to increase the network's ability to detect defective surfaces (Aslam et al., 2020; Wu and Lv, 2021).