Dense connection structures have garnered considerable interest in the deep learning domain due to their efficient feature reuse capabilities. However, their potential within Transformer architectures remains underexplored. To address the issues of feature collapse and inadequate multi-scale learning capabilities in classical Vision Transformers (ViT), we propose a feature learning method integrating dense connections, learnable weight matrices, and pyramid convolutions. The approach achieves efficient multi-layer feature reuse through dense connection structures while employing weight matrices to regulate the influence of different layers on final attention scores. We conducted experiments on four benchmark datasets: CIFAR-10, CIFAR-100, Oxford-Flowers-102, and Oxford-IIIT Pet. The consequences exhibit that the densely connected ViT exhibits superior performance in classification tasks and significantly improved multi-scale learning capabilities compared to baseline models.