Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Deno-IF: Unsupervised Noisy Visible and Infrared Image Fusion Method
Authors: Han Xu, Yuyang Li, Yunfei Deng, Jiayi Ma, Guangcan Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments validate the effectiveness and generalization of the proposed method for image fusion under various and variable noise conditions. The paper includes sections like "Experiments and Results", "Qualitative Results", "Quantitative Results", and performance tables (e.g., Table 1, Table 2). |
| Researcher Affiliation | Academia | 1 School of Automation, Southeast University, Nanjing, China 2 Electronic Information School, Wuhan University, Wuhan, China EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes mathematical formulations and iterative steps (e.g., equations 6a-6d for subproblems and subsequent solutions for Lt, Mt, Nt) but does not present them in a structured pseudocode or algorithm block format. |
| Open Source Code | Yes | The code is publicly available at https://github.com/hanna-xu/Deno-IF. |
| Open Datasets | Yes | We train Deno-IF on 2120 pairs of visible and infrared images across two datasets, including LLVIP [5] and M3FD [9]. |
| Dataset Splits | No | In the training phase, images are randomly cropped into patches of 128 128 for training. The evaluation is performed on image pairs introduced Gaussian or speckle noise with randomly distributed variance. The quantitative results are evaluated on 30 image pairs. This describes how training data is processed (cropped patches) and how evaluation is performed (on a set of image pairs) but does not specify a clear train/test/validation split of the main datasets (LLVIP, M3FD) in terms of percentages or absolute counts for the full dataset. |
| Hardware Specification | Yes | Experiments are conducted on an NVIDIA 3090 GPU. |
| Software Dependencies | No | In the joint denoising and fusion module, I2Former is updated with Adam Optimizer with batch size as 4 and epoch as 80. The paper mentions "Adam Optimizer" but does not specify its version or versions for other key software libraries like Python, PyTorch, or TensorFlow. |
| Experiment Setup | Yes | In the joint denoising and fusion module, I2Former is updated with Adam Optimizer with batch size as 4 and epoch as 80. Learning rate is 2e-4 with exponential decay. Hyper-parameters are λ = 1e3, η = 30. κ increases during training and equals the multiplication of 1e-6 and epoch. Numbers of blocks in Fig. 2 are Lr, L3 = 2, L1, L2 = 4. Experiments are conducted on an NVIDIA 3090 GPU. In convolutional low-rank optimization module, h and w are 128, c = 3. Kernel size of Ak( ) are k1, k2 = 12, k3 = 2. m, n = 256. We perform 30 iterations for each patch. For parameters in Eq. (5), β = 2. α, γ are empirically related to data characteristics. As an unsupervised method, for a patch x, we roughly estimate its noise level as n = E [| (x) (G(x))|], where G( ) is Gaussian blur. α is 200n and 80n for visible and infrared data respectively. β = 15n. |