Deno-IF

November 4, 2025 ยท View on GitHub

Code of "Deno-IF: Unsupervised Noisy Visible and Infrared Image Fusion Method" (NeurIPS 2025 Spotlight).
Paper

Introduction

This paper proposes an unsupervised noisy visible and infrared image fusion method, termed as Deno-IF. It consists of two modules:

First, based on the convolutional low-rank property of high-quality data, the convolutional low-rank optimization module decomposes clean component from the noisy input through convolution nuclear norm minimization in an unsupervised manner. The decomposed data provides the optimization guidance for the joint denoising and fusion module.

Then, the joint denoising and fusion module takes noisy source images as input and outputs the fused image. The network includes intra-modal recovery and inter-modal recovery and fusion, with self- and cross-modal attention to deal with complex individual-modal and complementary multi-modal information.

The framework of this method is shown below:


python=3.10
pytorch=1.13
pytorch-cuda=11.7
numpy=1.24.4
imageio=2.34.2
opencv-python=4.10
pandas=2.0.3
pillow=10.4
scikit-image=0.21
scipy=1.10.1

To train:

  • Prepare the training data:
    Put the training data in ./datasets/train/vis/ and ./datasets/train/ir/
  • Run python train.py

To Test:

  • Prepare the test data:
    Put the test data in ./datasets/test/vis/ and ./datasets/test/ir/
  • Choose the pretrained model:
    Set --ckpt in test.py as one of the following file path:
    i) LLVIP_M3FD.pth (for source images with relatively high resolution)
    ii) RoadScene_MSRS.pth (for source images with relatively low resolution).
  • Run python test.py