AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation
Authors: Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, Fahad Khan
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that the proposed method, Ada IR, achieves state-of-the-art performance on different image restoration tasks, including image denoising, dehazing, deraining, motion deblurring, and low-light image enhancement. The code is available at https://github.com/c-yn/Ada IR. |
| Researcher Affiliation | Collaboration | 1Technical University of Munich 2Inception Institute of Artificial Intelligence 3Mohammed Bin Zayed University of AI 4Australian National University 5University of Central Florida 6Linköping University |
| Pseudocode | No | The paper describes the FMi M and FMo M modules and their operations using mathematical formulas and descriptive text, along with diagrams (Fig. 3), but it does not present any formal pseudocode blocks or algorithms. |
| Open Source Code | Yes | The code is available at https://github.com/c-yn/Ada IR. |
| Open Datasets | Yes | For single-task image dehazing, we use one of the first standard dehazing datasets, SOTS (Li et al., 2018), which comprises 72,135 training images and 500 testing images. For single-task image deraining, we utilize Rain100L (Yang et al., 2019), which contains 200 clean-rainy image pairs for training and 100 pairs for testing. For single-task image denoising, we combine images of BSD400 (Arbelaez et al., 2010) and WED (Ma et al., 2016) datasets for model training; the BSD400 encompasses 400 training images, while the WED dataset consists of 4,744 images. Starting from these clean images of BSD400 (Arbelaez et al., 2010) and WED (Ma et al., 2016), we generate their corresponding noisy versions by adding Gaussian noise with varying levels (σ {15, 25, 50}). Denoising task evaluation is performed on BSD68 (Martin et al., 2001) and Urban100 (Huang et al., 2015). For motion deblurring, and low-light image enhancement, we used Go Pro (Nah et al., 2017) and LOL-v1 (Wei et al., 2018). We also evaluate on DPDD (Abuolaim & Brown, 2020) and AGAN (Qian et al., 2018) datasets, UAVDT (Du et al., 2018) and CSD (Chen et al., 2021d) dataset, Kodak24 (Rich, 1999). |
| Dataset Splits | Yes | For single-task image dehazing, we use one of the first standard dehazing datasets, SOTS (Li et al., 2018), which comprises 72,135 training images and 500 testing images. For single-task image deraining, we utilize Rain100L (Yang et al., 2019), which contains 200 clean-rainy image pairs for training and 100 pairs for testing. For single-task image denoising, we combine images of BSD400 (Arbelaez et al., 2010) and WED (Ma et al., 2016) datasets for model training; the BSD400 encompasses 400 training images, while the WED dataset consists of 4,744 images. Starting from these clean images of BSD400 (Arbelaez et al., 2010) and WED (Ma et al., 2016), we generate their corresponding noisy versions by adding Gaussian noise with varying levels (σ {15, 25, 50}). Denoising task evaluation is performed on BSD68 (Martin et al., 2001) and Urban100 (Huang et al., 2015). |
| Hardware Specification | Yes | All experiments are conducted on NVIDIA Tesla A100 40G GPUs using Py Torch. |
| Software Dependencies | No | The paper mentions 'using Py Torch' but does not specify a version number for PyTorch or any other software libraries or dependencies. |
| Experiment Setup | Yes | For training, we adopt a batch size of 32 in the all-in-one setting and a batch size of 8 in the single-task setting. The network optimization is achieved through an L1 loss function, employing the Adam optimizer (β1 = 0.9 and β2 = 0.999), with a learning rate of 2e 4, over the course of 150 epochs. During the training process, cropped patches sized at 128 128 pixels are provided as input, with additional augmentation applied via random horizontal and vertical flips. |