Residual Diffusion Deblurring Model for Single Image Defocus Deblurring
Authors: Haoxuan Feng, Haohui Zhou, Tian Ye, Sixiang Chen, Lei Zhu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our method achieves competitive performance on real-world defocus deblurring image datasets across both perceptual and distortion evaluation metrics. We conduct extensive experiments on DPDD dataset and demonstrate the state-of-the-art performance on it and enhance the perceptual quality significantly. Also, we evaluate our generalization ability on real-world outof-distribution datasets such as Real DOF dataset and CUHK dataset. |
| Researcher Affiliation | Academia | Haoxuan Feng1, Haohui Zhou1, Tian Ye1, Sixiang Chen1, Lei Zhu1,2* 1The Hong Kong University of Science and Technology (Guangzhou) 2The Hong Kong University of Science and Technology EMAIL, EMAIL |
| Pseudocode | Yes | The complete inference process is detailed in Algorithm 1. Algorithm 1: Inference procedure of our methods Require: Pre-deblur module pθ, defocus map estimator gθ, denoising network fθ, blurred image Ib, num of training timestep Tl and num of inference timestep Ti Ensure: Deblurred image I 1: Ip = pθ(Ib) Got Pre-deblurred image 2: Imap = gθ(Ip) Got defocus map 3: xt N(0, Id) Sampling noise from gaussian 4: for t in range (Tl, 1, Tl Ti ) do 5: zt N(0, Id) 6: xt 1 = mtxt + ktfθ(xt, t, c) + σβtzt 7: Reverse diffusion process 8: end for 9: return Ip + xt 1 Final deblurred image |
| Open Source Code | No | The paper does not contain an explicit statement about releasing source code for the methodology described, nor does it provide a direct link to a code repository. |
| Open Datasets | Yes | We train our methods on DPDD dataset (Abuolaim and Brown 2020), which includes 2,000 images distributed across 500 scenes. ... To assess the generalization capabilities of our models, we also tested them on the CUHK dataset (Shi, Xu, and Jia 2014) which consists of 1,000 blurred images from the Internet... Additionally, we evaluated our models on the Real DOF (Lee et al. 2021) dataset, which comprises 50 scenes with corresponding all-in-focus images captured by Sony a7R IV cameras. Detailed dataset statistics are available in Table 1. |
| Dataset Splits | No | The paper mentions training on the DPDD dataset and evaluating on the DPDD test dataset and Real DOF test set, but it does not specify the exact split percentages or sample counts for these datasets (e.g., how the DPDD dataset is divided into train/validation/test sets). |
| Hardware Specification | Yes | For evaluation of diffusion deblurring performance, we measure the PSNR, SSIM (Wang et al. 2004), LPIPS (Zhang et al. 2018) and FID192 (Heusel et al. 2018) between blurred image and ground truth images on a single Nvidia A800SXM4 GPU with a batch size of 1, indicating that the experiment is not affected by batch-dependent noise. |
| Software Dependencies | No | We implement our models using Py Torch (Paszke et al. 2019). No specific version of PyTorch or any other software dependencies with version numbers are provided. |
| Experiment Setup | Yes | We use Adam W (Loshchilov and Hutter 2017) with β1 = 0.9 and β2 = 0.999 and weight decay to 0.01 for training. The network is trained for 150k steps with an initial learning rate of 2e 4. Also, the cosine learning rate scheduler is to enhance training stability, returning to the initial learning rate after 200 step updates. For the diffusion hyperparameters, we choose DDIM (Song, Meng, and Ermon 2022) sampler and set the number of training steps and inference steps to 1,300 and 3, respectively. Data augmentation and loading involve using a crop size of 320 320 with a batch size of 8, and augment the data with horizontal flipping, vertical flipping and random rotation. |