Multi-modal knowledge graph completion (MMKGC) is essential for leveraging diverse data sources to enhance reasoning and fill missing information in real-world knowledge graphs. However, existing MMKGC methods still face significant challenges in dealing with modality noise and missing modalities, which hinder effective and robust multi-modal fusion. To address these issues, we propose a Denoiser-Reinforcement Learning Framework (DRLF) that enhances the fusion process by appropriately mitigating noise and dynamically adapting to missing information. The proposed framework comprises two core modules: (1) a Reward-Guided Representation Filter, which dynamically detects and suppresses noisy signals in modality inputs; and (2) a Reinforcement Learning-based Adaptive Modality Fusion Module, which employs an Actor-Critic architecture to adaptively optimize modality weights. Furthermore, DRLF introduces a relation-aware fine-tuning mechanism during the decoding stage to further improve relation modeling capabilities. Experimental results on two public MMKGC datasets show that DRLF consistently surpasses prevailing approaches across evaluation metrics such as MRR and Hit@K, and achieves performance on par with state-of-the-art models. Notably, it delivers more stable and robust results under challenging conditions with substantial modality missing and noise.