Compressive Image Recovery Using Recurrent Generative Model

Code corresponding to the paper : (https://arxiv.org/abs/1612.04229)
Forked from the original code for RIDE, which can be found here

Results

• Missing Pixel Interpolation

Original Image	Masked Image	During Gradient Ascent	Recovered Image

• Single Pixel Camera Reconstruction

Original Image	Initial Image	During Gradient Ascent	Recovered Image

Requirements

Same as original RIDE code (https://github.com/lucastheis/ride/)

Usage

1. Missing Pixel Interpolation

• Run python experiments/map_interpolate_stack.py for default parameters. Following options can be changed :

-m/--model  <Path to the model. Trained 1 layer and 2 layer models available in models/>
-d/--data   <Path to the test images in mat format. Images chosen from BSDS dataset in the paper available in data/>
-h/--holes  <Fraction of pixels removed from the image at random. Default is 70%>
-m/--momentum <Momentum set for gradient ascent in image reconstruction>
-l/--lr     <Learning Rate set for gradient ascent in image reconstruction>
-N/--niter  <Number of iterations for gradient ascent>
-p/--path   <Path to save the resulting images>
-q/--mode   <Mode to run Caffe in>
-D/-device  <Device ID for GPU>
-s/--size   <Size of test images>
-f/--flip   <Flag to carry out direction flipping as mentioned in paper>
-e/--ent_max <For thresholding posterior entropy as mentioned in paper>
-r/-resume   <For resuming the gradient ascent from previous npy file at certain iteration>
-I/--index   <To select which test image to work on from the mat file>

• The test image will be divided into four parts and each gradient ascent will run on each part simultaneously. To stitch the four reconstructed parts use python experiments/stitch_stack.py. Specify index of the test image using -I and iteration to choose for the reconstructed npy file using -i option

2. Single Pixel Camera Reconstruction

• Create compressive sensing matrix using python experiments/create_Phi.py

• Run python experiments/map_single_pixel_stack.py for default parameters. Following options can be changed :

-m/--model  <Path to the model. Trained 1 layer and 2 layer models available in models/>
-d/--data   <Path to the test images in mat format. Images chosen from BSDS dataset in the paper available in data/>
-n/--noise_std <For adding noise to the sensed measurements. By default no noise is added>
-d/--momentum <Momentum set for gradient ascent in image reconstruction>
-l/--lr     <Learning Rate set for gradient ascent in image reconstruction>
-N/--niter  <Number of iterations for gradient ascent>
-p/--path   <Path to save the resulting images>
-q/--mode   <Mode to run Caffe in>
-D/-device  <Device ID for GPU>
-s/--size   <Size of test images>
-f/--flip   <Flag to carry out direction flipping as mentioned in paper>
-e/--ent_max <For thresholding posterior entropy as mentioned in paper>
-r/-resume   <For resuming the gradient ascent from previous npy file at certain iteration>
-K/--image_num   <To select first K images from mat file to test >