Autofocus research based on an optical model accepted to ICCV 2025

Research Problem

Imaging systems with a shallow depth of field have a narrow range of sharp focus and are sensitive to small errors in focus position. Severe defocus further obscures structure and texture in the image, leaving less reliable information to guide focus adjustment. Conventional methods usually rely on repeated sampling and search. Methods that learn to predict defocus distance directly face a different problem. Images captured on opposite sides of the focal plane can look similarly blurred, making it difficult to determine which direction to move. This study addresses how to reduce the search required while maintaining reliable autofocus over a wide range of defocus distances.

Method Overview

We propose an autofocus method that combines image sharpness prediction with an optical model of the imaging system. A neural network estimates sharpness from a defocused image. The optical model then relates this estimate to defocus distance to guide the next focus adjustment.

The key contribution is the use of sharpness as an intermediate representation between image observations and focus adjustments. Combining predictions from image features with the physical constraints of the optical model helps guide the search for the correct focus position.

Key Findings

Across the industrial samples tested in the paper, our method covers a focusing range of approximately 35 times the depth of field, achieving performance at the current state of the art. It also achieves a focus success rate above 95% within five adjustments. The experiments include severely defocused starting positions. They show that the method can bring images into focus in fewer than five adjustments across a wide range of initial defocus distances.

Summary

This work addresses autofocus under shallow depth of field and severe defocus by combining an optical model with sharpness estimates learned from images. It offers a practical approach to autofocus in precision imaging and industrial vision.

The research has been accepted to the IEEE/CVF International Conference on Computer Vision (ICCV 2025). Read the full paper, Optical Model-Driven Sharpness Mapping for Autofocus in Small Depth-of-Field and Severe Defocus Scenarios.