CoRev2: A Unified Contrastive and Restorative Self-Supervised Learning Method With Masked Prompting Tuning for Surface Defect Inspection

This is the authors' abstract. We don't add key points for this paper.

Due to marked differences between natural images and defect images, the “supervised pretraining and full fine-tuning (FT)” paradigm has been proven to be suboptimal for defect inspection tasks. Recent advanced works attempt to mitigate the issue via self-supervised learning (SSL). However, the existing SSL works still have some issues: 1) the shortage of scale information in SSL hampers its representation capability and 2) the existing tedious and expensive FT strategy insufficiently exerts the potential of learned representation on downstream tasks. To address the above limitations, a unified contrastive and restorative SSL framework with masked prompting tuning (CoRev2) is proposed. CoRev2 aims to provide powerful and generalizable representation for diverse defect inspection scenarios. First, CoRev2 unites multiscale information, contrastive learning, and restorative learning to excavate required information in a complementary manner for defect inspection tasks. It will enhance the representation’s capability to identify complex defects. Subsequently, CoRev2 proposes a masked prompting tuning (MPT) strategy to achieve efficient knowledge transfer. The MPT strategy retrospects the pretrained knowledge by the prompt, and guides the learning process of new tasks with the cross-task information exchange mechanism. Experimental results on six challenging defect inspection datasets demonstrate that CoRev2 significantly surpasses other SSL methods. Meanwhile, CoRev2 can flexibly combine with various tailored defect inspection methods and bring further gains.