Learnable Blur Kernel for Single-Image Defocus Deblurring in the Wild
Authors: Jucai Zhai, Pengcheng Zeng, Chihao Ma, Jie Chen, Yong Zhao
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Competitive experimental results confirm that with a learnable blur kernel, the generated defocus map can achieve results comparable to supervised methods. In the single-image defocus deblurring task, the proposed method achieves state-of-the-art results, especially significant improvements in perceptual quality, where PSNR reaches 25.56 d B and LPIPS reaches 0.111. |
| Researcher Affiliation | Collaboration | 1 Shenzhen Graduate School, Peking University 2 Peng Cheng Laborator {jucaizhai, zpceng, machihao}@stu.pku.edu.cn, yongzhao@pkusz.edu.cn, chenj@pcl.ac.cn |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide a specific link or explicit statement about releasing its source code. It only mentions using code and weights provided by other authors for comparison. |
| Open Datasets | Yes | We use the dataset DPDD provided by (Abuolaim and Brown 2020a) for training and testing. This dataset has 500 sets of images, and each set of images includes a defocus blurred image, a pair of DP views, and an all-in-focus (Ai F) image with a resolution of 1680 1120. Here, like most methods (Abuolaim et al. 2021b), flowing the settings, 500 groups have been divided into 350, 74, and 76 groups according to the training set, validation set, and test set. We also use the CUHK dataset (Shi, Xu, and Jia 2015) and the Google Pixel DP dataset (Abuolaim and Brown 2020a) to verify the generalization of the network. |
| Dataset Splits | Yes | 500 groups have been divided into 350, 74, and 76 groups according to the training set, validation set, and test set. |
| Hardware Specification | Yes | We implemented the method using Pytorch and trained it on an NVIDIA RTX 3090 GPU. |
| Software Dependencies | No | The paper mentions 'Pytorch' but does not provide specific version numbers for it or any other software components, which are required for reproducibility. |
| Experiment Setup | Yes | The hyperparameter λ is set to 10 5, and the learning rate is set to 2 10 5. ... The 512 512 single-image and the defocus map are fed into the network. The number of iterations of the simulated annealing algorithm is set to 2 104, the hyperparameters α, β are set to 0.012 and 0.002, respectively. The initial learning rate is set to 2 10 4, which decreases by half every 30 epochs. ... The batch size of both networks is set to 4 and optimized using the Adam optimizer, where b1=0.9, b2=0.999. |