Solving Inverse Computational Imaging Problems using Deep Pixel-level Prior

Code for Single Pixel Camera reconstruction corresponding to the paper : (https://arxiv.org/abs/1802.09850) Derived from the original PixelCNN++ code which can be found here

Results

• Single Pixel Camera reconstruction with 25% measurement rate

Original Image	Masked Image	During Inference	Recovered Image

Requirements

• Tensorflow 1.2.0

• GPU with CUDA 8.0

Usage

1. A PixelCNN++ model trained on 64x64 ImageNet can be found here. Download the three files and save them in the saves folder.

2. For SPC reconstruction at 25% measurement rate with default parameters, run python reconstruct_spc.py