Contrastive Semantic‐Guided Image Smoothing Network

Image smoothing is a fundamental low-level vision task that aims to preserve salient structures of an image while removing insignificant details. Deep learning has been explored in deal with the complex entanglement semantic and trivial However, current methods neglect two important facts smoothing: 1) naive pixel-level regression supervised by limited number high-quality ground-truth could lead domain shift cause generalization problems towards real-world images; 2) texture appearance closely related object semantics, so requires awareness difference apply adaptive strengths. To address these issues, we propose novel Contrastive Semantic-Guided Smoothing Network (CSGIS-Net) combines both contrastive prior facilitate robust smoothing. The supervision signal augmented leveraging undesired effects as negative teachers, incorporating segmentation tasks encourage distinctiveness. realize proposed network, also enrich original VOC dataset enhancement labels, namely VOC-smooth, which first bridges segmentation. Extensive experiments demonstrate CSGIS-Net outperforms state-of-the-art algorithms large margin. Code are available at https://github.com/wangjie6866/CSGIS-Net.