FusionDN: A Unified Densely Connected Network for Image Fusion
Authors: Han Xu, Jiayi Ma, Zhuliang Le, Junjun Jiang, Xiaojie Guo12484-12491
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Qualitative and quantitative results demonstrate the advantages of Fusion DN compared with state-of-the-art methods in different fusion tasks. |
| Researcher Affiliation | Academia | 1Electronic Information School, Wuhan University, Wuhan 430072, China 2School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001, China 3College of Intelligence and Computing, Tianjin University, Tianjin 300350, China |
| Pseudocode | No | No pseudocode or algorithm blocks were found. |
| Open Source Code | Yes | The code is available at: https://github.com/hanna-xu/Fusion DN. |
| Open Datasets | Yes | The training and test sets are from three publicly available datasets: Road Scene for task11, the dataset provided by (Cai, Gu, and Zhang 2018)2 for task2, and Lytro Multi-focus3 for task3. Thereinto, Road Scene dataset is a new infrared and visible image dataset released by ourselves to remedy shortcomings in existing datasets. The new dataset has 221 aligned Vis and IR image pairs containing rich scenes such as roads, vehicles, pedestrians and so on. These image pairs are highly representative scenes from the FLIR video4. We preprocess the background thermal noise in the original IR images, accurately align the Vis and IR image pairs, and cut out the exact registration regions to form this dataset. It solves the problems in existing datasets such as few image pairs, low spatial resolution and extreme lack of detailed information in infrared images. Source images in the training datasets are cropped to patches of size 64 64. As for multi-focus images, due to the lack of aligned dataset, the source images are enlarged and flipped (either horizontally or vertically) to obtain more training data. λ is set as 12, 15 and 11, respectively. c is set as 13, 8 and 1 correspondingly. α is set as 5e-5 and β is set as 3e3. The parameters in Dense Net are updated by RMSProp Optimizer with the learning rate set as 1e-4. In the training phase, 177 Vis/IR pairs, 60 under/over-exposure pairs, and 10 far/near-focused pairs are used for training, respectively. In the testing phase, the numbers of image pairs in these 3 fusion tasks are 44, 30 and 10 correspondingly. 1https://github.com/hanna-xu/Road Scene 2http://rit-mcsl.org/fairchild//HDRPS/HDRthumbs.html 3https://mansournejati.ece.iut.ac.ir/content/lytro-multi-focusdataset 4https://www.flir.com/oem/adas/adas-dataset-form/ |
| Dataset Splits | No | In the training phase, 177 Vis/IR pairs, 60 under/over-exposure pairs, and 10 far/near-focused pairs are used for training, respectively. In the testing phase, the numbers of image pairs in these 3 fusion tasks are 44, 30 and 10 correspondingly. |
| Hardware Specification | No | No specific hardware details (like GPU or CPU models, or memory specifications) used for running experiments were provided. |
| Software Dependencies | No | The paper mentions 'RMSProp Optimizer' and 'VGG-16 network' but does not specify any software dependencies with version numbers. |
| Experiment Setup | Yes | λ is set as 12, 15 and 11, respectively. c is set as 13, 8 and 1 correspondingly. α is set as 5e-5 and β is set as 3e3. The parameters in Dense Net are updated by RMSProp Optimizer with the learning rate set as 1e-4. |