Balancing framework: Enhanced performance through contrastive masked encoders and gradient feature

Self-supervised learning (SSL) has shown strong potential for learning visual representations from unlabeled data, but transferring these representations effectively to object detection remains challenging. Existing SSL paradigms have complementary strengths and weaknesses: contrastive learning provides strong instance discrimination but lacks spatial sensitivity, while masked image modeling preserves spatial structure but is less effective at distinguishing instances. As a result, current methods often perform well on either classification or detection, but struggle to balance both. To address this problem, we propose Contrastive Masked Histogram-Decoupled Detector (CMHD), a dual-branch SSL framework that combines masked spatial reasoning with gradient-aware contrastive learning. The online branch improves spatial understanding by reconstructing masked regions, while the target branch enhances structural and discriminative representations through gradient-based features. We further introduce a random coordinate attention module to fuse cross-branch features and improve multi-scale representation learning. Extensive experiments on MS COCO, PASCAL VOC, ImageNet-1K, and other benchmarks show that CMHD achieves more balanced performance across classification and detection tasks, outperforming representative pure contrastive learning and masked image modeling baselines. Ablation studies further confirm the effectiveness of each component in the proposed framework.