SRUNet++: Skip connection & Residual blocks in U-Net++ for plug and play image restoration

Image restoration seeks to improve degraded images, recovering high-quality versions. However, existing model-based methods are limited by predefined priors, and suffered with suboptimal denoising performance, while deep plug-and-play denoisers often suffer from detail loss and weak noise generalization. To overcome these limitations, we propose SRUNet++, a plug-and-play image restoration network. Firstly, multi-scale features are aggregated via dense skip connections to integrate low-frequency semantics and high-frequency details. Secondly, a multi-layer residual series (MLRS) module is introduced to enhance local feature reuse and suppress noise interference. Finally, training with observed images and large-scale random noise level maps as input boosts the denoising and generalization of the model. Experimental results demonstrate that SRUNet++ significantly outperforms existing methods in both image denoising and deblurring tasks.