Task-Gated Multi-Expert Collaboration Network for Degraded Multi-Modal Image Fusion
Authors: Yiming Sun, Xin Li, Pengfei Zhu, Qinghua Hu, Dongwei Ren, Huiying Xu, Xinzhong Zhu
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that TG-ECNet significantly enhances fusion performance under diverse complex degradation conditions and improves robustness in downstream applications. The code is available at https: //github.com/Lee X54946/TG-ECNet. |
| Researcher Affiliation | Academia | 1School of Automation, Southeast University, Nanjing, China 2College of Intelligence and Computing, Tianjin University, Tianjin, China 3Engineering Research Center of City Intelligence and Digital Governance, Ministry of Education of the People s Republic of China, Tianjin, China 4Haihe Lab of ITAI, Tianjin, China 5School of Computer Science and Technology, Zhejiang Normal University, Jinhua, China. Correspondence to: Pengfei Zhu <EMAIL>. |
| Pseudocode | No | The paper describes the model architecture and mechanisms but does not include explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at https: //github.com/Lee X54946/TG-ECNet. |
| Open Datasets | Yes | We conduct experiments on both our constructed dataset and the EMS dataset (Yi et al., 2024). |
| Dataset Splits | Yes | Our De MMI-RF dataset includes 6 types of degradation: high/medium/low levels of Gaussian noise, haze, defocus blur, and striped noise. Typical cases of the dataset are shown in Fig. 4, which includes both ground and drone scenarios. De MMI-RF has 26631 training datasets and 9895 testing datasets, providing a powerful benchmark for degraded image fusion. |
| Hardware Specification | Yes | All experiments in this paper were performed on 6 NVIDIA Ge Force RTX 4090 GPUs |
| Software Dependencies | Yes | the model was implemented using the Py Torch 1.12.0 framework. |
| Experiment Setup | Yes | During the training phase, we used the Adam optimizer to optimize the network, setting the initial learning rate to 1.0 10 4 and adjusting it using the cosine annealing strategy. In addition, we randomly cropped the images to a size of 128 128 pixels for training. In each small batch, data augmentation was performed by flipping the images horizontally or vertically to expand the training sample size. We conduct experiments on both our constructed dataset and the EMS dataset (Yi et al., 2024). We trained a single model under 6 degradation settings. The first stage training process lasted for 30 epochs, and the model was directly tested across multiple restoration tasks. The second stage training process lasted for 30 epochs, and the model was directly tested across multiple restoration and fusion tasks. In experiments, the number of experts N and the number of selected experts K were heuristically set to 11 and 6, respectively. |