LMIF: A Large-scale Multi-focus Image Fusion Dataset Based on Smartphones

Official Repository for LMIF Dataset | 基于智能手机的大规模多焦点图像融合数据集

📖 Project Introduction

Due to the limited depth of field of optical lenses, a single photo cannot make all objects in the scene clear at the same time. Multi-focus image fusion technology solves this problem by fusing multiple images with different focal planes into an all-in-focus clear image.

Most existing public multi-focus datasets are collected by professional cameras, which cannot simulate the real shooting characteristics of mobile phones in daily scenarios. To fill this gap, we construct the LMIF dataset captured by mainstream consumer smartphones, which can better evaluate the generalization ability of fusion algorithms on real mobile photography data.

Paper Information

• Title: LMIF: A Large-scale Multi-Focus Image Fusion Dataset Based on Smartphones
• Conference: 2025 IEEE 3rd International Conference on Computer, Vision and Intelligent Technology (ICCVIT)
• Affiliation: College of Electronic and Information Engineering, Hebei University
• Corresponding Author: Shuaiqi Liu (shdkj-1918@163.com)

📦 Dataset Overview

Basic Information

• Total image pairs: 229 pairs (A:66 / B:63 / C:100)
• Two versions provided: Raw unprocessed data + Preprocessed standardized data (512×512 JPG)
• Shooting devices: 12 mobile phone models from 7 mainstream brands (Huawei, iPhone, OPPO, Honor, Redmi, iQOO, Samsung, vivo, Realme, OnePlus, Xiaomi)
• Shooting scenes: Campus, indoor desk, balcony, library and other daily real environments
• License: Non-commercial, only for academic research, algorithm verification and teaching demonstration

Preprocessing Pipeline

We processed raw mobile images to eliminate interference such as jitter, illumination difference and resolution mismatch:

1. Luminance equalization: SHINE color normalization toolbox to unify brightness and contrast
2. Image registration: SIFT feature matching to correct camera shake offset
3. Image cropping & resizing: Uniform 512×512 via nearest-neighbor interpolation

📥 Dataset Download

Full datasets (which can be used as image registration and fusion)

The download link for the full dataset is https://zenodo.org/records/21457298, for people in china, you can download by https://download.scidb.cn/download?fileId=9812e228a8f8f16ad73fa2cd5cfc07ba&path=/V1/LMIF.zip&fileName=LMIF.zip

Test datasets (which can be used as image fusion)

The download link for the test dataset (The preprocessed and registered image has a size of 512×512) is https://www.researchgate.net/publication/410637545