Recent research in emotion recognition in conversation (ERC) applied anchor-based contrastive learning to enhance the discrimination of similar emotions, but the failure to further optimize anchor positions has hindered the improvement of recognition accuracy. In this paper, we introduce a novel Optimized Anchor Contrastive Learning (OptiACL) framework to address the problem of similar emotion distinction in Multimodal ERC. Our method allows for fusing multimodal information via tensor ring decomposition to achieve intra-modal and inter-modal interactions. The fused representations are structurally regularized via label encoding anchors, dynamically optimized through interactive contrastive loss. To enhance the discrimination and confidence of anchors as classifiers, an anchor angle auxiliary loss and an adaptive threshold mining module are proposed for optimizing the positions of anchors. We also introduce an unsupervised consistency learning for noisy samples to align with refined anchors. Experimental results on two benchmarks indicate that our framework further improves the classification ability of emotion anchors.