Image Deblurring Principles and Practice - From Motion Blur to Defocus Recovery

Understanding Image Blur - The Degradation Model

Image blur is a quality degradation caused by camera or subject motion during exposure, lens defocus, or atmospheric turbulence. Mathematically, it is modeled as the convolution of a sharp image f with a blur kernel (PSF: Point Spread Function) h, plus additive noise n:

g = h * f + n

where g is the observed blurry image and * denotes convolution. The goal of deblurring is to recover the original sharp image f from g and h (when known).

Types of blur:

• Motion blur: Directional blur from camera shake or subject movement. PSF is line-shaped
• Defocus blur: Circular blur in out-of-focus regions. PSF is disk-shaped (pillbox)
• Atmospheric blur: Random blur from atmospheric turbulence. Problematic in astronomy and long-range imaging
• Spatially-varying blur: PSF differs across image positions. Includes rotational and zoom blur

Why deblurring is difficult: Deblurring is fundamentally an inverse problem with three core challenges: noise amplification, ringing artifacts, and solution non-uniqueness. With noise present, naive inverse filtering causes explosive noise amplification, producing unusable results without proper regularization.

Non-Blind Deblurring - Restoration with Known PSF

Non-blind deblurring recovers sharp images when the blur kernel (PSF) is known. Processing in the frequency domain is fundamental, with noise-robust regularization being the key to practical results.

Inverse Filter: The simplest approach, dividing by H(u,v) in frequency domain: F(u,v) = G(u,v) / H(u,v). However, noise is amplified to infinity at frequencies where H approaches zero, making this impractical for real images.

Wiener Filter: An optimal filter considering the noise-to-signal power spectrum ratio:

F(u,v) = [H*(u,v) / (|H(u,v)|² + K)] × G(u,v)

K represents the noise-to-signal ratio (NSR), typically set between 0.001-0.01. Larger K increases noise suppression but reduces restoration sharpness. Implementable in OpenCV using cv2.filter2D() combined with FFT operations.

Richardson-Lucy (RL) Deconvolution: An iterative method based on Poisson noise model, widely used in astronomy and microscopy. Updates the estimate each iteration, typically converging in 20-50 iterations. Naturally satisfies non-negativity constraints but over-iteration amplifies noise causing ringing artifacts.

Total Variation (TV) Regularization: Preserves edges while suppressing noise by minimizing ||g - h*f||² + λ×TV(f). The parameter λ (range 0.001-0.1) controls the balance between edge sharpness and noise suppression, providing excellent results for natural images.

Blind Deblurring - Simultaneous PSF Estimation and Image Recovery

Blind deblurring simultaneously estimates both the sharp image and PSF from a blurry observation without prior knowledge of the blur kernel. Since PSF is unknown in most real photography scenarios, this is the most practically important deblurring technique.

MAP (Maximum A Posteriori) estimation: The classical approach assumes priors on both image and PSF, maximizing posterior probability through alternating optimization:

• Image prior: Natural image gradients follow heavy-tailed distributions (hyper-Laplacian)
• PSF prior: Non-negative, sums to 1, sparse (most elements near zero)
• Alternating optimization: Fix f, estimate h → Fix h, estimate f → Repeat until convergence

Coarse-to-fine multi-scale strategy: Blind deblurring easily falls into local optima, so image pyramids are constructed starting estimation from the coarsest scale. PSF estimates at coarse scales initialize the next finer scale, progressively increasing resolution. Typically 4-6 pyramid scales are used for robust convergence.

Edge-based PSF estimation: The Cho-Lee (2009) method estimates PSF using only image edges (high-gradient regions). Edge regions strongly preserve blur direction and magnitude information, enabling efficient and accurate estimation. A two-stage approach using shock filters for edge enhancement followed by gradient-domain PSF estimation achieves 10x speedup over previous methods.

Deep Learning Deblurring - End-to-End Restoration

Deep learning deblurring methods learn direct mappings from blurry to sharp images without explicit PSF estimation. Rapidly advancing since 2017, these approaches significantly outperform traditional methods on standard benchmarks.

DeblurGAN (2018): GAN-based motion blur removal using a ResNet encoder-decoder generator trained with adversarial loss + perceptual loss. Achieves PSNR 28.7dB on the GoPro dataset (3,214 pairs). Inference speed is approximately 50ms for 720p on GPU.

DeblurGAN-v2 (2019): Introduces Feature Pyramid Network (FPN) for multi-scale feature utilization. A lightweight version using MobileNet-v2 backbone achieves 10x speedup while maintaining quality, enabling real-time processing on mobile devices.

MPRNet (2021): Multi-Stage Progressive Restoration Network restores images through 3 progressive stages. Each stage passes encoder-decoder output to the next, with residual learning for detail correction. Achieves PSNR 32.66dB on GoPro, setting state-of-the-art at the time of publication.

Restormer (2022): Applies Transformer architecture to image restoration using Multi-Dconv Head Transposed Attention for efficient global dependency capture at high resolutions. Achieves PSNR 32.92dB on GoPro and 31.22dB on HIDE dataset. Computation cost is approximately 300ms for 1280x720 on A100.

NAFNet (2022): Nonlinear Activation Free Network achieves Restormer-equivalent performance at half the computational cost through SimpleGate and Simplified Channel Attention innovations, reaching PSNR 33.69dB on GoPro.

Spatially-Varying Blur and Video Deblurring

Real-world photography commonly produces spatially-varying blur where blur direction and magnitude differ across the image. Video deblurring additionally leverages temporal information for superior restoration quality.

Causes of spatially-varying blur:

• Camera rotation: Roll rotation causes different blur directions at image center versus periphery
• Depth of field: Defocus blur size varies with distance from the focal plane
• Object motion: Different objects moving at different velocities within the scene
• Zoom blur: Radial blur pattern from zoom operation during exposure

Handling spatially-varying blur: The basic approach divides images into patches (64x64 to 128x128) and estimates local PSF per patch, with continuity constraints between adjacent patches. Deep learning methods use Deformable Convolution for spatially-adaptive filtering, which has become the dominant approach for handling non-uniform blur.

Video deblurring: Leveraging consecutive frames enables higher quality restoration than single-frame methods. EDVR (Enhanced Deformable Video Restoration) aligns and fuses information from 5 adjacent frames using Deformable Convolution, improving PSNR by 1-2dB over single-frame approaches.

Event camera fusion: Event cameras (DVS) with microsecond temporal resolution record the blur formation process. Using event data during conventional camera exposure enables recovery of severe motion blur previously considered impossible. E-CIR achieves PSNR above 34dB using event data as restoration guidance.

Specialized books on image restoration are available on Amazon

Practical Deblurring Tools and Quality Assessment

This section covers concrete tools, parameter settings, and quality evaluation methods for applying deblurring in production workflows, including common failure patterns and their solutions.

Desktop tools:

• Topaz Sharpen AI: Deep learning commercial tool with automatic detection of motion blur, defocus, and soft focus modes. Processing time approximately 5 seconds for 24MP on GPU
• Adobe Photoshop: Filter > Sharpen > Shake Reduction. Auto-estimates PSF with Wiener filter restoration. Manual PSF direction and length adjustment available
• DxO PhotoLab: Lens profile-based optical correction using lens-specific PSF databases for high-accuracy restoration

Python libraries:

• scikit-image: Provides skimage.restoration.wiener() and richardson_lucy()
• OpenCV: Wiener filter implementable via cv2.filter2D() + FFT
• DeepDeblur: PyTorch-based deep learning deblurring with pre-trained models ready for immediate use

Quality metrics:

• PSNR: Basic reference metric. Above 30dB is good, above 35dB is excellent
• SSIM: Structural similarity. Above 0.9 is good quality
• LPIPS: Perceptual similarity (lower is better). Below 0.1 is good
• NIQE: No-reference quality metric measuring deviation from natural image statistics

Common failures and solutions: Ringing artifacts (ripples near edges) are mitigated by adjusting regularization parameters. Noise amplification is addressed by applying denoising first or using joint deblur-denoise methods. Halo effects from over-sharpening are avoided by using conservative processing strength settings.