Multi-Focus Image Fusion via Explicit Defocus Blur Modelling
Authors: Yuhui Quan, Xi Wan, Zitao Tang, Jinxiu Liang, Hui Ji
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on benchmark datasets demonstrated the effectiveness of our approach. Code https://github.com/Tangzitao/DMANet Extended version https: //github.com/Tangzitao/DMANet/raw/main/paper.pdf Experimental results on benchmark datasets have demonstrated the superior performance of our DMANet over stateof-the-art techniques. Quantitative comparison Table 1 presents the quantitative results on MFI-WHU which includes GTs. Our DMANet achieves the best ARank and excels across most metrics, particularly showcasing a significantly higher PSNR compared to other methods. Tables 2 to 5 display the results on Lytro, MFFW, SIMIF, and OR-PAM, respectively, where GTs are unavailable. Our DMANet consistently ranks first in terms of ARank across all datasets, demonstrating its robust performance. Ablation Studies We constructed several baseline methods to evaluate the contributions of different components in our approach: (a) 1scale: only one scale is used the DBE, with the Conv LSTM replaced with a convolutional block with a similar size; (b) w/o DBE: remove the DBE from DMANet and the DME accepts the source images as input. (c) w/o SR: discard the selfrecurrent (SR) structure by replacing the Con LSTM in SRM with a convolutional block of a similar size. (d) w/o Ldefocus: remove the defocus estimation loss. Table 6 lists these baselines results on the Lytro dataset, where all baselines show a noticeable performance decrease. These observations confirm that each component in our approach contributes noticeably to the performance. To analyze the effectiveness of UAF, we decrease its threshold γ from 0.1 to 0. Note that when γ = 0, the fusion becomes using binary masks without uncertainty awareness The corresponding PSNR results on the MFI-WHU dataset are shown in Table 7. As γ decays, the PSNR performance decreases, demonstrating the effectiveness of UAF. |
| Researcher Affiliation | Academia | 1School of Computer Science and Engineering, South China University of Technology, China 2School of Computer Science, Peking University, China 3Department of Mathematics, National University of Singapore, Singapore EMAIL, EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology and model architecture using text and diagrams (e.g., Figure 2) but does not include a distinct pseudocode or algorithm block. |
| Open Source Code | Yes | Code https://github.com/Tangzitao/DMANet |
| Open Datasets | Yes | Datasets Following Wang et al. (2023), model training is done on 10000 synthesized image pairs sourced from Agustsson and Timofte (2017) and Wang et al. (2019). The performance is then evaluated on five benchmark datasets: MFIWHU (Zhang et al. 2021a), Lytro (Nejati, Samavi, and Shirani 2015), SIMIF (Tsai 2024), MFFW (Xu et al. 2020b), and OR-PAM (Zhou et al. 2022). |
| Dataset Splits | No | The paper mentions using 10000 synthesized image pairs for training and evaluates on five benchmark datasets, but does not explicitly provide specific training/test/validation splits (e.g., percentages, sample counts, or references to predefined splits for these datasets). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running the experiments. It only mentions computational complexity in terms of model size and inference time. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | No | The paper mentions the training loss components (Lfusion, Ldefocus) and a weight parameter (beta) and initialization for learnable parameters (σn = 0.5(n 1)), but it lacks specific hyperparameters such as learning rate, batch size, number of epochs, or the specific optimizer used for training the deep learning model. |