Multi-label Recognition under Noisy Supervision: A Confusion Mixture Modeling Approach

Abstract

Multi-label recognition is a critical task in artificial intelligence, aiming to identify every object present in an image. Designing a multi-label classifier is a nontrivial task both from data collection and modeling perspectives. Collecting multiple labels for each image is extremely time-consuming and costly, which often leads to noisy annotations. Furthermore, a robust classifier that performs reliably well in the presence of such noisy labels demands meticulous modeling and learning criterion design. In this work, we propose a novel probabilistic confusion model that effectively incorporates inter-label interactions in causing label noise. The proposed model incorporates a latent variable, building a hierarchical structure to the label noise generation, and represents the label noise as a mixture of confusions caused by various classes. Under the proposed multi-label confusion mixture (MCM) model, we design an end-to-end learning criterion along with a sparsity regularization, that effectively estimates the true multi-label classifier. Experiments with various real-world datasets showcase the effectiveness of our approach.