Generating Real-World Data-Guided Synthetic Noisy Labels for Robust Image Classification

The core message of this article is to propose a new algorithm that leverages real-world information to inject controllable noise into clean datasets, generating real-data guided synthetic noise (RGN) with characteristics of being less time-consuming, easy-to-generate, and close to real-world scenarios. This algorithm can be applied to any dataset in any noisy scenario to construct suitable and normative datasets with noisy labels for robust algorithm testing.