The article presents ‘Swintormer’, a new model for image deblurring that offers improved performance and lower memory usage. It uses a diffusion model to generate latent prior features, helping to restore detailed images. The model also employs a sliding window strategy in specialized Transformer blocks for efficient inference. Compared to the current top-performing GRL method, Swintormer drastically decreases computational complexity and improves the Signal-to-Noise Ratio (SNR) for defocus deblurring. This allows higher resolution images to be processed on devices with limited memory.
Publication date: 12 Jan 2024
Project Page: https://github.com/bnm6900030/swintormer
Paper: https://arxiv.org/pdf/2401.05907
