Deep Partial Label Learning Algorithm Based on Joint Optimization Strategy

Abstract

Partial label learning (PLL) is a novel weakly supervised learning framework, where each training sample corresponds to a set of candidate labels and the real label of that sample is hidden in the set. Most of the current researches on PLL algorithms are based on traditional machine learning algorithms, resulting in failure to extract deep features of images. On the other hand, those algorithms using deep learning technology have no effective disambiguation strategy, which leads to the inability to distinguish labels. To overcome these two shortcomings, this paper proposes a deep PLL algorithm called PL-DJO (Partial label-Deep Joint Optimization). We use ResNet34 as a feature extractor, define a loss function for partial label data, which disambiguates the set of candidate labels to obtain pseudo-real labels. Then, a joint optimization strategy is used to iteratively update model parameters, and the pseudo-real labels are updated by predicted labels, which reduces the noise rate of pseudo-real labels and significantly improves the model performance. Experiments on four real and four synthetic partially labeled datasets demonstrate the superiority of the PL-DJO algorithm compared with other state-of-the-art methods.

Keywords: Image Classification and Recognition, Deep Partial Label Learning, Joint Optimization Strategy, ResNet34