Weakly supervised watermark removal using multi-view visual models and inpainting

Digital images widely adopt watermarking techniques for copyright protection and source attribution. However, effectively and accurately removing watermarks without pixel-level annotations remains a significant challenge. This study proposes an integrated approach that combines weakly supervised learning with advanced image inpainting techniques, aiming to achieve precise watermark detection and removal without relying on complex mask annotations. By applying diverse preprocessing strategies, the original images are transformed into multiple visual representations and processed through several deep learning models for training and prediction. An ensemble voting mechanism is then used to enhance the stability and accuracy of watermark detection. For the detected watermark regions, multi-model feature response maps are fused to determine the areas that require restoration, which are subsequently processed by an image inpainting model for masking and visual reconstruction. This method effectively integrates multi-view feature learning with image inpainting, reducing the cost of annotation while improving the naturalness and consistency of watermark removal. In the future, it can be extended to more application scenarios and handle more complex and diverse watermark patterns, demonstrating strong potential for practical use.