Objects of different scales exhibit varied characteristics within images, necessitating network architectures capable of handling multi-scale information for optimal performance in complex scenarios. However, ResNet struggles with detecting small or occluded objects. In this work, we introduce the Multi-Dimensional Attention- Reparameterized Multi-Branch Network (MDA-RepXNet), a novel multi-scale network designed to enhance performance in challenging conditions. MDA-RepXNet employs a Structurally Reparameterized Multi-Branch Network (RepXNet) as its backbone, incorporating multi-scale branches to manage different feature scales and thus improving small object recognition. Additionally, we propose a Multi-Dimensional Attention (MDA) mechanism to enhance feature fusion. This design significantly improves image feature extraction and object detection while maintaining high computational efficiency. Experimental results demonstrate that our approach increases the mean Average Precision (mAP) by 3.9% on the MS COCO dataset compared to the baseline, underscoring the effectiveness and superiority of the proposed method.