Progress and Thinking on Self-Supervised Learning Methods in Computer Vision: A Review

Deep learning (DL) methods have been widely studied and applied in the field of computer vision (CV) over the past decades. The biggest disadvantage of classic DL methods is that they strongly rely on a large number of labeled samples. In engineering application, however, it is much expensive and even impossible to generate so many high-quality labeled samples. For this purpose, self-supervised learning (SSL) methods have become a research hot spot in CV in recent years due to their strong ability of learning representation without manually labeled images. So far, SSL has made strides in CV, but it is far from maturity and still faces some underlying challenges. The main purpose of this article is to review the latest development of SSL methods and applications, summarize key technologies and challenges, and discuss the trends. First, the development history of SSL methods in CV is outlined. Then, the existing SSL methods in CV are classified into four main categories and typical applications SSL in CV are summarized. Finally, key technologies of SSL are refined and future trends are discussed. This article can help researchers to quickly understand the current progress of SSL in CV.