Deep Self-Supervised Learning for Oracle Bone Inscriptions Features Representation

In this paper, we design a two-branch deep learning framework to tackle the problem of self-supervised representation learning for Oracle Bone Inscriptions (OBIs). This problem is very complicated in that, unlike natural-photos, OBI images present more abstract content and suffer from different drawing styles, resulting in the failure of many existing self-supervised learning methods to describe them accurately. The core idea of our framework is that we design two OBI-specific pretext tasks, i.e. rotation and deformation. These two kinds of pretext tasks can provide strong supervision signals for OBI features learning. And we perform OBI recognition downstream task to evaluate our self-supervised learned features. Experimental results show that, under the same dataset, our proposed method outperforms jigsaw and matting based self-supervised learning methods.