This paper introduces a perceptive self-supervised learning network (PSLNet) for removing watermarks from noisy images. Traditional methods often rely on supervised learning with reference images, which are difficult to obtain in real-world scenarios. PSLNet addresses this issue by using a parallel network architecture that includes a self-supervised learning mechanism (SSL) and two sub-networks: a denoising and watermark removal network (DWRN) and a stacked perception network (SPN). The SSL mechanism generates paired watermark images through a self-supervised approach, while the DWRN and SPN handle the actual watermark removal. The DWRN consists of two improved U-Nets, one for denoising and the other for watermark removal, operating in parallel. The SPN, implemented using a VGG network, extracts texture information to enhance the performance of the watermark removal process. A mixed loss function, combining structural and texture losses, is used to improve the robustness and effectiveness of PSLNet. Experimental results on various datasets demonstrate that PSLNet outperforms existing methods in terms of quantitative metrics such as PSNR, SSIM, and LPIPS, as well as qualitative visual quality. The proposed method is also shown to be efficient in terms of computational complexity.This paper introduces a perceptive self-supervised learning network (PSLNet) for removing watermarks from noisy images. Traditional methods often rely on supervised learning with reference images, which are difficult to obtain in real-world scenarios. PSLNet addresses this issue by using a parallel network architecture that includes a self-supervised learning mechanism (SSL) and two sub-networks: a denoising and watermark removal network (DWRN) and a stacked perception network (SPN). The SSL mechanism generates paired watermark images through a self-supervised approach, while the DWRN and SPN handle the actual watermark removal. The DWRN consists of two improved U-Nets, one for denoising and the other for watermark removal, operating in parallel. The SPN, implemented using a VGG network, extracts texture information to enhance the performance of the watermark removal process. A mixed loss function, combining structural and texture losses, is used to improve the robustness and effectiveness of PSLNet. Experimental results on various datasets demonstrate that PSLNet outperforms existing methods in terms of quantitative metrics such as PSNR, SSIM, and LPIPS, as well as qualitative visual quality. The proposed method is also shown to be efficient in terms of computational complexity.