Deep learning has achieved significant advancements in the area of object detection. However, in scenarios where there is limited access to annotated samples due to factors like scarce data resources and high labeling costs, traditional fully supervised learning methods face significant challenges. In semi-supervised object detection (SSOD), training error from pseudo labels accumulate and impair model performance. We mitigate this by incorporating confidence scores into the loss function to minimize error propagation. And at the same time, the model also converges faster. To address labeled-data scarcity, we introduce the intelligent augmentation module by segmenting images into foreground and background, and applying a diffusion model to inpaint the background regions to increase the number of labeled data. Experimental validation on VOC and LEVIR dataset demonstrate the effectiveness of the proposed method. In particular, under the condition of 1% labeled data, our method achieves an improvement of 4.93% and 5.80% in mAP@0.5 over the baseline on these two datasets, respectively.