Attention mechanisms have become critical elements in deep learning models, especially for tasks such as Image Captioning (IC) and Visual Question Answering (VQA). However, current attention modules predominantly focus on spatial dependencies, often overlooking the critical aspect of multi-dimensional perception fusion, which is vital for achieving a deep and comprehensive semantic understanding. To address this limitation, we propose a novel design, namely Multi-Dimensional Fusion Transformer Network (MDFNet). MDFNet integrates Swin Transformers with shifted window partitioning for feature map division and spatial/channel-wise attention, effectively aggregating global and local features for enhanced multi-modal understanding and accurate scene interpretation. This method enables the model to more effectively grasp complex relationships within the data, leading to enhanced multi-modal reasoning and a more accurate interpretation of complex scenes. Extensive experiments demonstrate the effectiveness of MDFNet, achieving a new state-of-the-art performance with a CIDEr score of 135.2 on the COCO Karpathy test split. In addition to its superior performance in image captioning, MDFNet also enhances the model’s overall robustness, representing a significant advancement in the evolution of attention mechanisms for multimedia tasks.