Micro-expression recognition (MER) is challenging because extracting locally subtle changes in micro-expressions (MEs) is extremely difficult. Optical flow describes the variations between frames, which can effectively suppress facial identity information while characterizing ME movements well. Thus, several optical flow-based methods have been proposed to recognize MEs. However, these approaches using architectures with one branch for one input or multiple branches for multiple inputs do not reveal discriminative features, which leads to inferior performance. This paper proposes a novel homo-modal framework based on optical flow for the MER problem, termed homo-modal attention refinement network with Distance Correlation (HARN-DC). Concretely, HARN-DC consists of three components, i.e., an expression feature learning module, an expression-dilated feature learning module, and a classification branch. First, two identical structural Inception networks with a channel-wise attention module are designed to learn parallel global and local expression features based on the same ME’s optical flow images. Second, to expand ME representations, a dilated loss incorporating Distance Correlation is proposed to amplify the differences between the two branches’ features. Last, the emotion categories are predicted via a fusion of expression-dilated features in the classification branch. Extensive experiments conducted on the composite database published by MEGC 2019 validate the effectiveness of HARN-DC under leave-one-subject-out (LOSO) cross-validation and composite database evaluation (CDE) protocol. The results indicate that our proposed approach can generate discriminative features and yield promising performance gains. Moreover, the results also show that HARN-DC is competitive with comparable state-of-the-art methods on MER.