Attention-based Semi-supervised Partial Label Learning on Web Image Classification

Partial label learning s a type of weakly supervised learning, which learns to predict one label as the correct answer from a given candidate label set. It is a great challenge to improve the accuracy of recognition because of the combination of the two difficult learning conditions, partial supervised learning and partial supervised learning. However, these methods all use similar matrices for tag propagation. We introduced the attention mechanism, proposed a attention-based semi-supervised partial label learning (ASPLL) method to address the label contamination issue in PLL through reliable label propagation. Using attention mechanism instead of similarity matrix can significantly improve the accuracy of the algorithm. We evaluate the performance of our method on real-world web image datasets. The experimental results on the web image dataset show that in the semi supervised partial labeling web image classification, ASSPL algorithm is significantly better than the mainstream semi-supervised partial labeling learning algorithms, such as SSPL, PARM.