Digital image watermarking, while crucial for copyright protection, often compromises image integrity and background information. Traditional watermark removal methods suffer from inaccurate segmentation and inconsistent reconstruction across watermark regions. To address these challenges, we propose a novel watermark removal framework that integrates diffusion models with knowledge distillation, forming a three-stage pipeline: precise localization, progressive reconstruction, and semantic alignment. First, we introduce a boundary-aware segmentation network to improve the accuracy of watermark mask prediction, especially around complex edges. Then, a diffusion-based restoration strategy is applied, where Gaussian noise is injected into the masked regions and progressively refined to recover plausible background content. Finally, to overcome style inconsistency in diffusion-generated regions, we design a knowledge distillation scheme that transfers semantic features from a teacher model trained on clean images to guide the reconstruction process. This alignment encourages both semantic and stylistic consistency between the restored region and the original background. Extensive experiments on synthetic and real-world watermark datasets demonstrate that our method achieves superior performance in terms of removal accuracy, reconstruction quality, and visual coherence, outperforming existing approaches.