IRNeXt: Rethinking Convolutional Network Design for Image Restoration
Authors: Yuning Cui, Wenqi Ren, Sining Yang, Xiaochun Cao, Alois Knoll
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments demonstrate that IRNe Xt delivers state-of-the-art performance among numerous datasets on a range of image restoration tasks with low computational complexity, including image dehazing, single-image defocus/motion deblurring, image deraining, and image desnowing. To demonstrate the effectiveness of our model, we evaluate IRNe Xt on 13 datasets for five image restoration tasks: single-image defocus deblurring, image dehazing, image deraining, image desnowing, and image motion deblurring. |
| Researcher Affiliation | Academia | 1School of Computation, Information and Technology, Technical University of Munich, Munich, Germany 2School of Ocean Information Engineering, Jimei Univeristy, Xiamen, China 3School of Cyber Science and Technology, Shenzhen Campus of Sun Yatsen University, Shenzhen, China. |
| Pseudocode | No | The paper describes the architecture and operations using diagrams and mathematical equations, but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | https://github.com/c-yn/IRNe Xt. |
| Open Datasets | Yes | We conduct single-image defocus deblurring experiments on the widely used DPDD (Abuolaim & Brown, 2020) dataset. We use a synthetic dataset RESIDE (Li et al., 2018a) and two real-world datasets, Dense-Haze (Ancuti et al., 2019) and NH-HAZE (Ancuti et al., 2020), to evaluate IRNe Xt. Following previous methods (Zamir et al., 2021; Tu et al., 2022), we train IRNe Xt on a composite dataset and compute the metrics on the Y channel in YCb Cr color space. Table 3 shows the results on Rain100L (Yang et al., 2017), Rain100H (Yang et al., 2017), and Test100 (Zhang et al., 2019b). We compare desnowing performance on three datasets: CSD (Chen et al., 2021c), SRRS (Chen et al., 2020), and Snow100K (Liu et al., 2018a). Following previous methods (Cho et al., 2021; Zamir et al., 2022; Tu et al., 2022; Wang et al., 2022), we train our model on the Go Pro (Nah et al., 2017) dataset. |
| Dataset Splits | Yes | DPDD is split into training, validation, and testing subsets with 350, 74, and 76 scenes, respectively. |
| Hardware Specification | Yes | All models are trained and evaluated on an NVIDIA Tesla V100 GPU. |
| Software Dependencies | No | We train our model using the Adam optimizer (Kingma & Ba, 2014) with the initial learning rate as 1e 4, which is gradually reduced to 1e 6 with cosine annealing (Loshchilov & Hutter, 2016). The inference time is tested in a synchronized manner by using torch.cuda.synchronize. (No specific versions for Python, PyTorch, CUDA are mentioned). |
| Experiment Setup | Yes | In all experiments, unless specified otherwise, the following hyper-parameters are used. We choose G = 8 and K = 3 in Eq. 7. We train our model using the Adam optimizer (Kingma & Ba, 2014) with the initial learning rate as 1e 4, which is gradually reduced to 1e 6 with cosine annealing (Loshchilov & Hutter, 2016). For data augmentation, we use random horizontal flips. With the exception of Go Pro (Nah et al., 2017), where n in Figure 3 (c) is set as 13, we set n = 15 for deraining and deblurring tasks, and n = 3 for dehazing and desnowing datasets. The model is trained for 3000 epochs. |