A PSO-optimized self-supervised convolutional network for robust image watermark removal.

With advancements in big data and internet technologies, images are becoming central to professional and recreational activities. While watermarks are crucial for copyright protection, their robustness against common manipulations such as compression or filtering also makes their removal for legitimate restoration purposes highly challenging. Traditional approaches to watermark removal almost always require manual tuning of parameters and complex optimization, resulting in limited applicability in practical use. To address these issues, this paper presents an improved self-supervised convolutional neural network (CNN) for image watermark removal, enhanced with Particle Swarm Optimization for hyperparameter tuning. The proposed approach performs self-supervised learning to generate training data, while avoiding the preparation of clean reference images, and benefits from PSO-based optimization in pursuing faster convergence and higher restoration quality. Experimental results on the PASCAL VOC 2012 dataset show that the proposed PSO-SWCNN has achieved the best performance among several state-of-the-art methods with the highest SSIM and PSNR. This integrated approach offers an effective and efficient solution in removing watermarks from images and restoring them for research and archival purposes.