Multimodal sentiment analysis (MSA) seeks to interpret human emotions by integrating text, images, and audio. Despite progress, existing methods struggle with class imbalance and ineffective cross-modal sentiment representation. We propose EMSA, an ensemble-based framework combining a primary and auxiliary model, both enhanced with image captioning and consistency loss. Caption generation via multimodal large language models strengthens visual-textual alignment, while consistency loss improves modality-specific sentiment learning. By training on both balanced and imbalanced datasets, EMSA mitigates class imbalance without distorting data distribution. Experiments on public benchmarks show that EMSA achieves state-of-the-art results, especially under class imbalance and complex cross-modal conditions.