Label Distribution Learning (LDL) is an emerging machine learning paradigm that uses label distributions instead of logical labels to effectively reduce information loss during the learning process. Applying LDL requires that the label distributions are explicitly given in the training set. However, obtaining the label distributions is costly and difficult to quantify. To address this challenge, many studies have focused on label enhancement, which aims to recover label distributions from logical labels. However, existing label enhancement methods are only applicable to datasets with balanced label information, and the recovery is not satisfactory in datasets containing label irrelevant features. To tackle these limitations, we propose a novel label enhancement method called Imbalanced Label Enhancement Based on Variational Information Bottleneck (VIB- ILE). Specifically, VIB-ILE mines the most relevant features for recovering the label distributions based on the variational information bottleneck and learns the joint implicit representations of these features with the logical labels. To improve the quality of the implicit representations, we employ a consistency regularization method. We recover label distributions based on the implicit representation. Moreover, we propose an asymmetric imbalanced label enhancement loss function for datasets with imbalanced label information. Finally, extensive experiments on real datasets prove the effectiveness of our method.