julian-becker / lensless

Reconstruct images from lensless coding cameras

lensless

Reconstruct images from lensless coding cameras

2D deconvolution using proximal gradient descent

The script realtime_deconv.m takes care of loading in measured data, preparing file names etc., then calls diffuser_2d_deconv_v2.m to do the deconvolution. diffuser_2d_deconv_v2 relies on the code in antipa/proxMin to function, so you'll need that. At the top of realtime_deconv are:

1. input_folder where you'll need to specify the path to the image to process
2. camera_type string containing camera type. Currently needs to be either 'pco' or 'flea3'. This must be set because each camera's raw data contains a different bias, which, if not properly subtracted, leads to poor reconstruction.
3. process_color String specifying which color channel to use. For flea3, mono is the only option, but for pco, choices are:

• 'mono': average RGB data in PSF and raw measurement
• 'red': use only red channel (after demosaicing)
• 'green': green only
• 'blue': blue
• 'all': solve all 3 colors independently, then create RGB output

At the top of diffuser_2d_deconv_v2, you can specify ds to be an integer. The images will all be downsampled (using a 'box' antialias filter) by this amount.

You can also set the regularizer parameters in the code, but this needs to be made more user friendly before it is recommended.