Li, X., Wu, S., Yuan, X., Xie, S., & Agaian, S. (2024). Hierarchical wavelet-guided diffusion model for single image deblurring. The Visual Computer. https://doi.org/10.1007/s00371-024-03631-9
Abstract:
Image deblurring is a critical preprocessing step in various high-level vision tasks, including facial recognition, medical imaging, and object detection. Blurred artifacts can arise from multiple factors, such as camera shakes and fast-moving objects. Recently, diffusion models (DMs) have made significant progress in image deblurring by employing a sequence of denoising refinements conditioned on the blurry input. However, existing DM-based methods often neglect the potential of incorporating frequency information, limiting their ability to reconstruct fine textures crucial for high perceptual quality. To this end, we propose a hierarchical wavelet-guided diffusion model (HWDM) for single image deblurring. HWDM integrates multi-level frequency information from wavelet-transformed domains into the denoising network of DM, facilitating the restoration of high-quality deblurred images. Specifically, HWDM consists of three components: primary frequency recovery network (PFRNet), which aims to restore essential frequency information missing in the blurry image; multi-frequency extractor, which extracts multi-frequency information at various scales from PFRNet’s output using multi-level wavelet transforms; and DM’s denoising network (DMDNet), into which the frequency information is hierarchically integrated via a cross-attention mechanism, effectively utilizing fine-grained and multi-scale information. Extensive experiments on synthetic and real-world blur datasets demonstrate that HWDM outperforms state-of-the-art methods in perceptual quality, producing more realistic and visually appealing deblurred images. This technology has enormous potential for applications in road traffic, medical imaging, remote sensing satellites.
License type:
Publisher Copyright
Funding Info:
There was no specific funding for the research done