We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Spatial-guided informative semantic joint transformer for single-image deraining.
- Authors
Li, Haiyan; Peng, Shaolin; Lang, Xun; Ye, Shuhua; Li, Hongsong
- Abstract
Single-image deraining is of great significance for image recognition and analysis. However, the majority of current methods face challenges such as incomplete removal of tiny rain streaks, blurred restoration of background structure, and feature interference. To address these issues, a spatial-guided informative semantic joint transformer (SISTrans) is proposed. Specifically, a high-dimensional spatial feature mapping module is put forward to restrain the receptive field of filters by increasing the spatial resolution of the image, guiding the entire module to focus on local features and thereby learn the distribution of small rain streaks. Subsequently, a wavelet-content-aware-based dual-level module is designed to capture the high-level semantic information by using an improved Swin transformer and to establish effective long-range dependencies via a shifted window mechanism, thereby enhancing the quality of background restoration. Ultimately, a dynamic hybrid cross-fusion module is proposed to effectively avoid feature interference by recalibrating the features of two branches and fusing the calibrated features with a set of learnable parameters. Extensive experiments conducted on eight commonly benchmark datasets demonstrate that the proposed SISTrans outperforms the state-of-the-art methods. Code is available at: https://github.com/SL-Pen/SISTrans.
- Subjects
TRANSFORMER models; IMAGE recognition (Computer vision); IMAGE analysis; SPATIAL filters; SPATIAL resolution; RAINFALL
- Publication
Journal of Supercomputing, 2024, Vol 80, Issue 5, p6522
- ISSN
0920-8542
- Publication type
Article
- DOI
10.1007/s11227-023-05697-z