Multimodal Data Dynamic Compression Algorithm Based on Semantic Importance

With the exponential growth of multimodal data in IoT, efficiently compressing data while preserving semantic integrity has become a critical challenge. However, due to the limitations of existing single-modality and non-dynamic approaches, semantic compression still suffers from low compression efficiency and incomplete semantic preservation. To address these limitations, this paper proposes a multimodal data compression algorithm based on dynamic semantic importance. Specifically, we improve the traditional text importance measurement algorithm TF-IDF by introducing a feature gradient-driven dynamic semantic importance measurement method. Combined with task-oriented weight allocation, it precisely quantifies the contribution of semantic features to task objectives. Furthermore, an compression threshold optimization algorithm is designed to dynamically balance semantic fidelity and resource constraints during compression. Simulation results demonstrate that the proposed compression method achieves a mean squared error (MSE) of 0.12 while compressing data volume by 50%. Under AWGN channels (SNR > 10 dB), low-threshold models (w0 < 0.4) maintain stable MSE below 0.12 with downstream flood detection accuracy reaching 95.8%.