Object detection models traditionally require large-scale labeled datasets, which increases both annotation cost and computational burden. To address this limitation, this paper proposes a self-supervised framework that integrates Vision Transformers (ViTs) with the Detection Transformer (DETR) architecture to enhance detection performance in low-label settings. We pretrain ViTs using DINO, MAE, and SimCLR strategies to enable robust feature extraction and contextual understanding, particularly for small and occluded objects. The proposed model—DETR with MAE-pretrained ViT—achieves a 5.2% improvement in mean Average Precision (mAP) and a 6.3% increase in Intersection over Union (IoU) on the COCO dataset compared to CNN-based DETR models. An ablation study demonstrates the effectiveness of contrastive learning for feature discrimination and masked autoencoders for spatial localization. Additionally, Grad-CAM visualizations confirm the interpretability and attention focus of ViT-based detectors. Despite higher computational requirements, the approach significantly reduces dependency on annotated data and improves generalization. Future work will focus on optimizing hybrid transformer backbones for real-time deployment in resource-constrained environments.