A multi-stage active learning framework with an instance-based sample selection algorithm for steel surface defect

Abstract

The application of deep learning (DL) for high-precision inspection to identify and locate the positions of each type of steel surface defect has demonstrated considerable potential for the quality control of steel products. However, the time-consuming and labor-intensive nature of manually labeling large amounts of data has limited DL’s broader deployment in this field. While traditional active learning methods can select the most valuable labels based on the amount of information, they fail to consider the positional and categorical information during the information computation process, thereby preventing the extraction of spatial information with multiple defects simultaneously. To address this challenge, this paper proposes a multi-stage active learning framework with an instance-based sample selection algorithm (MALF) for steel surface defects. Firstly, a soft weighted label assignment with prior information is constructed with the objective of achieving stable training and high-precision instance detection with a minimal amount of label annotation. Furthermore, when provided with high-precision instances, an independent evidence branch utilizing a reweighted Dirichlet distribution is capable of generating epistemic uncertainty with remarkable efficiency. Besides, a methodology based on diversity has been devised to ascertain the similarity with instance data as a diversity criterion, thereby obtaining detailed spatial information in multi-defect images. The results of experiments conducted on a variety of benchmark methods indicate that MALF is capable of filtering out more informative images for annotation while achieving higher accuracy with the same sample size.