<p>Graph contrastive learning (GCL) has gained significant research interest in node representation learning, due to its capability of effectively addressing the challenges posed by the scarcity or difficulty in obtaining labeled graph data. However, the encoder in GCL usually employs the aggregation-then-propagation mechanism, whose success on obtaining discriminative representation depends on the homophily of a graph. In this paper, considering that the inherent level of homophily constrains the efficacy of neighborhood information aggregation, we explicitly leverage homophily to refine the graph input, thereby injecting more structural and feature information into the graph. Then, with the homophily-boosted graph, a novel multi-level graph contrastive learning framework is proposed to make GCL benefit from hierarchical information at node, local, and global levels simultaneously. Meanwhile, we introduce two regularization terms to effectively mitigate local dimensional collapse maybe aroused by homophily-boosting. Finally, extensive experiments on five widely-used benchmark datasets are conducted to demonstrate the superiority of our proposed method. The code is available at <a href="https://github.com/S-su-code/h-GCL">https://github.com/S-su-code/h-GCL</a>.</p>