A multi-view decoupled fusion (MDF) framework was proposed to address the limitations of existing methods for multimodal emotion recognition in conversation (MERC) regarding context modeling and multimodal fusion. A multi-view context extraction module was designed to capture emotional dependencies in conversation through multi-scale contextual views from near to far. Furthermore, the discriminative ability was enhanced by combining supervised and unsupervised learning strategies. In addition, a multi-view modal decoupled fusion module was introduced, in which a Top-K mask mechanism was utilized to filter noise in attention matrices, and the expressive capability was improved by modeling the importance of each attention head. Experimental results on mainstream datasets demonstrated that, compared with the state-of-the-art method AdaIGN, the MDF achieved significant improvements in both context extraction and multimodal fusion. Specifically, on the IEMOCAP dataset, the accuracy (Acc) and weighted F1-score (W-F1) were improved by 2.38 and 2.31 percentage points, respectively; on the MELD dataset, Acc and W-F1 were improved by 0.44 and 0.13 percentage points, respectively. These results validated the effectiveness of the MDF framework.