Image Captioning with Multi-Scale Dilated Attention Mechanism

To address the limitations of traditional CNN-LSTM image captioning models in capturing key information and modeling long-range semantic dependencies, this study proposes an enhanced model named DLM that integrates a Multi-Scale Dilated Attention (MSDA) mechanism. The proposed DLM framework adopts DenseNet161 as the encoder backbone, integrating a multi-branch Multi-Scale Dilated Attention module to hierarchically fuse multi-resolution attention maps. Through a dilation rate hierarchy strategy and sparse feature fusion, the framework enhances the joint modeling capacity between local visual details and global semantic contexts. Experiments on the Flickr8K Dataset demonstrate that the proposed model achieves a 2.75% improvement in the BLEU-4 metric compared to the baseline model (DenseNet161+LSTM). Ablation experiments further validate the effectiveness of MSDA in complex scenarios, where hierarchical attention allocation significantly optimizes cross-modal alignment accuracy between images and text. This study provides a precision-and-efficiency balanced solution for image captioning tasks and offers valuable insights for the practical application of multimodal understanding technologies.