This paper proposes a perceptive self-supervised learning network (PSLNet) for noisy image watermark removal. Traditional methods often rely on reference images and degradation models to learn watermark removal, but obtaining reference images is difficult and collected images are often noisy. PSLNet addresses these issues by using a parallel network to remove both noise and watermarks. The upper network uses task decomposition to sequentially remove noise and watermarks, while the lower network uses a degradation model to simultaneously remove both. Paired watermark images are obtained through self-supervised learning, and paired noisy images are obtained through supervised learning. The two sub-networks interact and fuse obtained clean images to enhance structural information and pixel enhancement. A mixed loss is used to improve the model's robustness in terms of texture information. Comprehensive experiments show that PSLNet outperforms popular convolutional neural networks (CNNs) in noisy image watermark removal. The network includes a self-supervised learning mechanism, a denoising and watermark removal network (DWRN), and a stacked perception network (SPN). The DWRN consists of two parallel networks: an upper network for sequential noise and watermark removal, and a lower network for simultaneous removal. The SPN uses a 16-layer perception network based on VGG to extract texture information. A mixed loss function is used to combine structural and texture information for training. The network is evaluated on a dataset of 477 natural images from PASCAL VOC 2021, with 252 watermarked images for testing. The results show that PSLNet achieves higher PSNR, SSIM, and LPIPS values compared to other methods, demonstrating its effectiveness in noisy image watermark removal. The network is also efficient in terms of computational complexity, with fewer parameters than other methods. The results indicate that PSLNet is very effective for noisy image watermark removal.This paper proposes a perceptive self-supervised learning network (PSLNet) for noisy image watermark removal. Traditional methods often rely on reference images and degradation models to learn watermark removal, but obtaining reference images is difficult and collected images are often noisy. PSLNet addresses these issues by using a parallel network to remove both noise and watermarks. The upper network uses task decomposition to sequentially remove noise and watermarks, while the lower network uses a degradation model to simultaneously remove both. Paired watermark images are obtained through self-supervised learning, and paired noisy images are obtained through supervised learning. The two sub-networks interact and fuse obtained clean images to enhance structural information and pixel enhancement. A mixed loss is used to improve the model's robustness in terms of texture information. Comprehensive experiments show that PSLNet outperforms popular convolutional neural networks (CNNs) in noisy image watermark removal. The network includes a self-supervised learning mechanism, a denoising and watermark removal network (DWRN), and a stacked perception network (SPN). The DWRN consists of two parallel networks: an upper network for sequential noise and watermark removal, and a lower network for simultaneous removal. The SPN uses a 16-layer perception network based on VGG to extract texture information. A mixed loss function is used to combine structural and texture information for training. The network is evaluated on a dataset of 477 natural images from PASCAL VOC 2021, with 252 watermarked images for testing. The results show that PSLNet achieves higher PSNR, SSIM, and LPIPS values compared to other methods, demonstrating its effectiveness in noisy image watermark removal. The network is also efficient in terms of computational complexity, with fewer parameters than other methods. The results indicate that PSLNet is very effective for noisy image watermark removal.