Deep-learning-driven Cross-modal Image Fusion For Cartoon Captioning

Deep learning has revolutionized cross-modal understanding through its ability to process heterogeneous visual-textual data, which achieves encouraging performance in the image captioning domain. However, existing methods struggle with redundant cross-modal representations and autoregressive generation bottlenecks, leading to semantically misaligned captions. To address these challenges, an innovative data-driven cartoon image captioning framework (IBA-CD) is proposed through two synergistic advancements. First, IBA-CD develops an information bottleneck-driven cross-modal alignment mechanism that fusing principles from information theory and deep learning, to optimize semantic distillation. Such a mechanism suppresses redundant visual-textual mutual information while maximizing task-relevant correlations through variational inference. Second, IBA-CD pioneers a cascaded diffusion generation paradigm that reimagines text synthesis through bidirectional Transformer-based denoising processes, establishing non-autoregressive generation with multi-stage visual-semantic refinement. IBA-CD achieves component synergy through a bidirectional closedloop mechanism: The information bottleneck alignment dynamically injects distilled compact semantic features as conditional guidance into each denoising step of the diffusion network, enabling fine-grained visual-textual alignment through cross-modal attention mechanisms. Simultaneously, quality feedback from the generation process proactively optimizes the feature alignment intensity, forming an iterative refinement cycle that evolves from semantic compression to generation correction. This collaborative framework ultimately accomplishes efficient and precise cross-modal reasoning through tightly coupled visual-semantic distillation and progressive generative enhancement. Extensive experiments on benchmark datasets verify significant improvements in the cartoon image captioning task.