Research map / Image generation / Image restoration

Prior and Prediction Inverse Kernel Transformer for Single Image Defocus Deblurring

AAAI 2024 · Image generation › Image restoration · approach: Image deblurring · 8 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes P2IKT, a divide-and-conquer end-to-end defocus deblurring network that combines a Wiener-based inverse Gaussian kernel module and a learned inverse kernel prediction module, fused by a scale-recurrent transformer.

Problem

Single image defocus deblurring is hard because defocus blur has spatially varying sizes and shapes; end-to-end learned methods generalize poorly across cameras, while kernel-prior methods cannot handle irregular blur.

Approach

Treat defocus blur as either Gaussian-approximable or irregular: an inverse Gaussian kernel module (Wiener deconvolution with multi-size Gaussian kernels) handles the former, a kernel prediction network predicts an inverse kernel for the latter, and a scale-recurrent transformer predicts coefficient maps that adaptively fuse both outputs in a coarse-to-fine scheme.

Key results

• On DPDD, P2IKT reaches 26.29 dB PSNR vs 25.98 dB for Restormer, with 3.32M vs 26.10M parameters.
• On the RTF generalization test set, P2IKT achieves 25.85 dB PSNR, 0.839 SSIM and 0.207 LPIPS, best across metrics.
• Adding P2IKB (IGKM+IKPM) to SRT raises DPDD PSNR from 26.03 to 26.29 dB while adding only about 0.05M parameters.

Measured effects reported in the paper

Surprising findings

• Restormer, despite strong DPDD results, performs worse than IFAN and GKMNet on the low-resolution DED and RTF sets, and P2IKT has worse SSIM and LPIPS than Restormer on DPDD.

Limitations

• Fails to deblur cases in the DED dataset, as the authors acknowledge.
• Generalization across cameras is evaluated mainly on PSNR; SSIM and LPIPS are worse than Restormer on DPDD.
• Some baseline results were quoted from papers rather than re-run under identical conditions.

Experimental setup

Datasets: DPDD

Benchmarks: DPDD RealDoF LF-DOF DED RTF CUHK PixelDP

Metrics: PSNR SSIM LPIPS

Baselines: JNB EBDB DMENet DPDNet-S KPAC IFAN GKMNet DRBNet Restormer

Builds on: Scale recurrent scheme (Tao et al. 2018) GKMNet multi-size Gaussian kernel approximation Kernel prediction networks Restormer Wiener deconvolution

Closest papers by meaning

• Neumann Network With Recursive Kernels for Single Image Defocus Deblurring (CVPR 2023 · similarity 0.92)
• Residual Diffusion Deblurring Model for Single Image Defocus Deblurring (AAAI 2025 · similarity 0.91)
• DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.89)
• Parameterized Blur Kernel Prior Learning for Local Motion Deblurring (CVPR 2025 · similarity 0.88)
• Quad-Pixel Image Defocus Deblurring: A New Benchmark and Model (CVPR 2025 · similarity 0.87)
• Blur2Blur: Blur Conversion for Unsupervised Image Deblurring on Unknown Domains (CVPR 2024 · similarity 0.87)