SMD-Net: Multimodal change detection for content-adaptive semantic video transmission

The surge of video-centric applications has resulted in an overwhelming growth of data traffic, posing major challenges for efficient video transmission under bandwidth and latency constraints. Conventional video coding standards, though effective at exploiting signal-level redundancies, primarily optimize for pixel fidelity and often fail to preserve the semantic information most crucial for human or machine-centric tasks, particularly at low bitrates. To address these limitations, this paper proposes a novel Video Semantic Communication (VSC) framework that fundamentally shifts the optimization objective from pixel-level accuracy to task-driven semantic effectiveness. Our framework adaptively selects and transmits only semantically critical keyframes by leveraging a newly proposed Semantic Multimodal Difference Network (SMD-Net), which fuses visual and contextual textual information through an information-theoretic bottleneck principle. SMD-Net generates detailed pixel-wise semantic change maps and a semantic change score, enabling the content-adaptive selection of keyframes and efficient semantic encoding. Experimental results on video datasets demonstrate that the proposed method not only significantly reduces transmission overhead but also preserves task-relevant semantic content, yielding substantial gains in both human and machine-centric evaluation metrics compared to conventional approaches.