Realistic Compound-Lens Defocus Blur Synthesis

2026-02-10 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

POSTECH researchers introduce a novel pipeline for synthesizing realistic defocus deblurring datasets tailored for diverse compound lenses. This framework integrates efficient wave-optics Point Spread Function (PSF) computation using Debye Chirp Z-Transform (CZT) propagation, depth-aware defocus rendering with occlusion handling, and blur synthesis in a radiometrically linear space with camera Image Signal Processor (ISP) simulation. The pipeline generates CLDefocus, a large-scale synthetic dataset comprising 40,000 training, 1,000 validation, and 1,000 test pairs at 384x384 resolution, utilizing 700 distinct lens designs. Experiments demonstrate that models trained on CLDefocus achieve superior cross-device generalization compared to those trained on existing real (DPDD) and synthetic (SYNDOF) datasets. The Debye CZT method also proves significantly faster and more stable than the Rayleigh-Sommerfeld integral for PSF computation.

Key takeaway

For Machine Learning Engineers developing image deblurring solutions, recognize that models trained on real-captured datasets may overfit to specific camera systems and be biased by ground-truth inconsistencies. To achieve robust cross-device generalization, prioritize training with physically grounded synthetic datasets like CLDefocus. Your models will benefit from diverse lens characteristics and photorealistic blur synthesis, leading to perceptually sharper restorations and improved performance on unseen camera types, including smartphones.

Key insights

Realistic synthetic defocus datasets improve deblurring model generalization across diverse camera lenses.

Principles

• Deblurring model performance is fundamentally constrained by training data realism and diversity.
• Physically grounded PSF computation and ISP-aware blur synthesis are crucial for photorealistic synthetic data.
• Real-captured datasets often contain inconsistencies that bias full-reference evaluation metrics.

Method

The pipeline computes PSFs via Debye CZT, quantizes depth into layers, performs occlusion-aware rendering in linear space, and applies ISP simulation for noise and saturation.

In practice

• Employ Debye CZT for efficient and stable wave-optics PSF generation.
• Perform blur synthesis in linear RGB space, then apply camera ISP for realism.
• Quantize scene depth using signed CoC values for efficient layered rendering.

Topics

• Defocus Deblurring
• Synthetic Data Generation
• Point Spread Function
• Wave Optics
• Debye CZT
• Compound Lenses

Code references

• lykelee/CLDefocus
• jax-ml/jax

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Realistic Compound-Lens Defocus Blur Synthesis — Synthesizing diverse, realistic defocus blur datasets improves deblurring model generalization across various camera systems.
• How this AI Learned to Refocus Any Photo — AI refocusing allows dynamic post-capture adjustment of photographic focus and depth of field from a single image.
• RTE-FM-Dehazer: Radiative Transfer Equation Inspired Flow Matching for Real-World Image Dehazing — RTE-FM-Dehazer combines the Radiative Transfer Equation with flow matching and a large synthetic dataset to achieve superior real-world image dehazing.
• Dehaze-GaussianImage: Zero-Shot Dehazing via Efficient 2D Gaussian Splatting Representation — Dehaze-GaussianImage uses 2D Gaussian Splatting for zero-shot dehazing, modeling haze as evolvable anisotropic Gaussian fields.
• Synthetic-to-Real Translation for Class-Agnostic Motion Prediction — Synthetic-to-real motion prediction for autonomous driving is significantly improved by integrating objectness priors and a specialized 4D LiDAR dataset.
• Rendering-Aware Bayesian 3D Gaussian Splatting with Native Uncertainty and Adaptive Complexity Control — Bayesian 3DGS provides native uncertainty and adaptive complexity for novel-view synthesis, outperforming standard methods in active view selection.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.