In the sequential recommendation task, the recommender generally learns multiple embeddings from a user's historical behaviors, to catch the diverse interests of the user. Nevertheless, the existing approaches just extract each interest independently for the corresponding sub-sequence while ignoring the global correlation of the entire interaction sequence, which may fail to capture the user's inherent preference for the potential interests generalization and unavoidably make the recommended items homogeneous with the historical behaviors. In this paper, we propose a novel Dual-Scale Interest Extraction framework (DSIE) to precisely estimate the user's current interests. Specifically, DSIE explicitly models the user's inherent preference with contrastive learning by attending over his/her entire interaction sequence at the global scale and catches the user's diverse interests in a fine granularity at the local scale. Moreover, we develop a novel interest aggregation module to integrate the multi-interests according to the inherent preference to generate the user's current interests for the next-item prediction. Experiments conducted on three real-world benchmark datasets demonstrate that DSIE outperforms the state-of-the-art models in terms of recommendation preciseness and novelty.