Code for reproducing the result of paper [Blind Image Deconvolution by Generative-based Kernel Prior and Initializer via Latent Encoding]

• PyTorch >= 1.10.0
• Requirements: opencv-python, tqdm

Download the pretrained models of kernel Generator(netG) and Initializer(netE) from Google Drive to the models folder.

Put the test images in the ./datasets/Lai/uniform folder. Reproduce results reported in the paper.

python deblur_lai.py

The generated N blur kernels are combined into a three-dimensional array of shape (N, kernel_size, kernel_size), and stored as an npz file in the ./datasets/kernel folder.

To train the kernel initializer, a clean image dataset is used for convolving with the blur kernels to produce blurred images. We used the OpenImage dataset and place the folder open_val into ./datasets. Other clean image datasets could also be considered.

python train_generator.py --kernel_size [size of kernel] --kernel_path [path to kernel] --save_path [path to save model]

python train_initializer.py --kernel_size [size of kernel] --kernel_path [path to kernel] --save_path [path to save model]

If our work is useful for your research, please cite our paper: