<p>When dealing with complex scenes containing multi-oriented objects, oriented object detection models can effectively capture object orientation information, but they typically require large amounts of accurately annotated rotated data for training. Semi-Supervised Oriented Object Detection (SSOOD) significantly reduces the dependence on annotated rotated data by leveraging unlabeled data while improving detection performance in complex scenes. Building on existing SSOOD methods, this paper proposes an innovative model termed CBLC-SOOD. The core contributions of our method are twofold. First, a module is designed to filter out noisy pseudo-labels and extract valuable information from background labels, thereby significantly improving the quality of dense pseudo-labels and providing more reliable supervision for model training. Second, a loss function based on contrastive background modeling is introduced, which establishes contrastive relationships between foreground and background regions from the raw outputs of the teacher and student models. This enhances the model’s ability to distinguish foreground objects from background areas, improving its robustness in complex scenes. To thoroughly assess the effectiveness of the proposed method, we conducted systematic experiments on the DOTA dataset using 10%, 20%, and 30% of the labeled data. Extensive experimental results demonstrate that CBLC-SOOD consistently outperforms baseline method with limited labeled data, exhibiting superior adaptability and robustness.</p>