The knowledge graph completion task aims to mine the missing fact triples in the knowledge graph based on the existing fact triples (head entity, relation, tail entity). Existing research work mainly focuses on using the structure in the knowledge graph. However, these works ignore that other modal information contained in the knowledge graph may also be effective for the knowledge graph completion task, such as the description text of the entity and the image matching the entity. Moreover, since task-specific knowledge is usually not injected into the general pre-training model, how to fuse task-related knowledge in the process of extracting modal information becomes crucial. Additionally, since different modal features contribute differently to knowledge graph completion, how to effectively retain useful multimodal information is also a great challenge. To solve the above problems, this paper proposes a multimodal knowledge graph completion method that integrates task knowledge. We use a fine-tuned multimodal encoder on the current task to obtain entity vector representations in different modalities. Then, we use a modal fusion-filtering module based on a recurrent neural network to remove task-independent multimodal features. Finally, we employ the simple isomorphic graph network to represent and update all features, thus effectively achieving multimodal knowledge graph completion. Experimental results show that our method can effectively extract information from different modalities, and can enhance the representation ability of entities through further multimodal filtering and fusion, thereby improving the performance of multimodal knowledge graph completion task.