DnCNN-B: a lightweight watermark image correction model based on DnCNN and bicubic interpolation

Digital watermarking safeguards digital products by embedding copyright information within the host data. Accurate recognition of extracted watermark images is crucial to the effectiveness of watermarking algorithms. Existing studies primarily focus on embedding and extraction processes, which are often complex and lack optimization for extraction quality. To address this issue, this study proposes a deep learning-based correction model to improve watermark recognition. The model employs a pre-trained denoising convolutional neural network (DnCNN) to enhance low-quality watermark images based on extraction outcomes. Subsequently, bicubic interpolation is applied for super-resolution processing to further improve image quality. The combined watermark optimization model is referred to as DnCNN-B. Experimental results demonstrate that the model effectively corrects errors in binary, grayscale, and color images, as well as QR codes and other watermark carriers. It also significantly enhances watermark recognition across various watermarking algorithms, embedding strengths, and attack levels. This approach improves watermark recognition without compromising the fundamental performance of the watermarking algorithm, thereby enhancing its practical applicability.