• Color Grading & Restoration
  • Blog
  • ABOUT
  • Contact
Menu

Zachary Halberd

  • Color Grading & Restoration
  • Blog
  • ABOUT
  • Contact

Deconvolutional Upscaling for Legacy Media

February 9, 2021

Case Study: Shelter Island

Feature documentary upscale

Source Media: 1280x720p - Mpeg4

Target Deliverable: 2560x1440p - ProRes 422 HQ

Director Michael Canzoniero at MCM Creative in NYC called me and asked if it was possible to clean up and upscale a feature documentary called Shelter Island prior to an Amazon streaming distribution sale. Unfortunately, the original media was no longer available and the master had already gone through various unknown encodes.

The source media was a compressed file created over ten years ago (2010-2011) via an unknown method from an unknown source. The film was initially recorded in 1280x720 with the HVX200 in DVCPRO HD on P2 cards. Initial analysis indicated significant intermediate compression due to the presence of heavy discrete cosine transform artifacts. MCM’s request was to remove the compression artifacts, camera noise, and upscale to 1080p or greater. while ensuring that the color management was within modern standards. We also did a quality control pass in Davinci Resolve to ensure there were no colors or levels out of range.

 

We recommend setting the playback settings to 4k for the video below:

 

Resolving the finer details

Original source media 720p at 20x magnification

Original source media 720p at 20x magnification

Deconvolved image 1440p at 20x magnification

Deconvolved image 1440p at 20x magnification

 

Removing compression artifacts

Original source media 720p at 10x magnification

Original source media 720p at 10x magnification

Deconvolved image 1440p at 10x magnification

Deconvolved image 1440p at 10x magnification

 

Deconvolving finer structure

Good edge detection and deconvolution which doesn’t impart too much posterization or deconvolving hoops.

Original source media 720p at 10x magnification

Original source media 720p at 10x magnification

Deconvolved image 1440p at 10x magnification

Deconvolved image 1440p at 10x magnification

 
Original source media 720p - no magnification

Original source media 720p - no magnification

Edge detection at 5x magnification

Edge detection at 5x magnification

Deconvolved edge detection at 5x magnification

Deconvolved edge detection at 5x magnification

Original source media 720p at 10x magnification

Original source media 720p at 10x magnification

Deconvolved image 1440p at 10x magnification

Deconvolved image 1440p at 10x magnification

 

A closer look at resolving detail and compression noise reduction. The algorithm that I used for this project is particularly effective with geometrically uniform shapes like the car in the background. Although the algorithm renders a bit too much uniformity to the “smoothed” polygons, a slight blend or even randomized grain could help bring some life back into the image.

Original source media 720p at 20x magnification

Original source media 720p at 20x magnification

Deconvolved image 1440p at 20x magnification

Deconvolved image 1440p at 20x magnification

Original source media 720p at 10x magnification

Original source media 720p at 10x magnification

Deconvolved image 1440p at 10x magnification

Deconvolved image 1440p at 10x magnification

 

As you can see in the following image set, it is easier to resolve the text on the coffee cup in the deconvolved image. This is also another good example of how the denoising/deblocking process can impart a slightly smooth characteristic to the image. Although the noise in the initial image is not pleasing, there is a nice “life” to it that is missed. I ultimately chose to render a slight film grain over the top of the export in order to compensate.

Original source media 720p with no magnification

Original source media 720p with no magnification

Original source media 720p at 10x magnification

Original source media 720p at 10x magnification

Deconvolved image 1440p at 10x magnification

Deconvolved image 1440p at 10x magnification

Edge detection with original source media 720p at 10x magnification

Edge detection with original source media 720p at 10x magnification

Edge detection with deconvolved media 1440p at 10x magnification

Edge detection with deconvolved media 1440p at 10x magnification

← VapourSynth Installation for Python 3.6.8Superior Deinterlacing and Upscaling With Machine Learning →

Latest Posts

Featured
DOC_WORKFLOW_20240428.png
May 12, 2024
Documentary Finishing Workflow Demystified
May 12, 2024
Read more →
May 12, 2024
Using AI Models to Colorize Monochromatic Archival
September 30, 2022
Using AI Models to Colorize Monochromatic Archival
September 30, 2022
Read more →
September 30, 2022
Archival Restoration - Experimenting with Python Using VapourSynth + Resolve
October 25, 2021
Archival Restoration - Experimenting with Python Using VapourSynth + Resolve
October 25, 2021

Take a peak under the hood of my process to restore a heavily compressed archival film using the open source Python script-based image processing tool Vapoursynth.

Read more →
October 25, 2021
Qualitative Camera Test v2 - Summer 2021
October 25, 2021
Qualitative Camera Test v2 - Summer 2021
October 25, 2021

I convinced ten cinematographers to bring their cinema cameras to a qualitative camera test. Here are the results.

Read more →
October 25, 2021
axel-vazquez-nph6DcLTckI-unsplash.jpg
July 3, 2021
Premiere to Resolve and Back Again - A Remote Workflow Guide
July 3, 2021

For the love of God, please stop using shuttle drives. Just use the Resolve media manager to shrink ray your Premiere timeline down to a tiny portable file and send to your colorist!!! In this article, I show you how!

Read more →
July 3, 2021
VapourSynth Installation for Python 3.6.8
June 14, 2021
VapourSynth Installation for Python 3.6.8
June 14, 2021

My ongoing experiments with incorporating the Python-based image processing software, VapourSynth, into my color management and restoration pipeline has resulted in compiling a stable package which Windows users may use for further experimentation.

Read more →
June 14, 2021
Deconvolutional Upscaling for Legacy Media
February 9, 2021
Deconvolutional Upscaling for Legacy Media
February 9, 2021

MCM Creative in NYC reached out to Magic Lantern and asked if it was possible to clean up and upscale his feature doc Shelter Island prior to an Amazon streaming distribution sale. Unfortunately, the original media was no longer available and the master had already gone through various unknown encodes.

Read more →
February 9, 2021
Superior Deinterlacing and Upscaling With Machine Learning
February 2, 2021
Superior Deinterlacing and Upscaling With Machine Learning
February 2, 2021

Typically, filmmakers use the deinterlacing tool provided for free within their NLE or Davinci Resolve. This article will demonstrate why this approach can be destructive to the image and will offer a better alternative.

Read more →
February 2, 2021
Rendering With JPEG Compression(ism)
January 28, 2021
Rendering With JPEG Compression(ism)
January 28, 2021

A collection of photos taken with a cell phone at full magnification yields interesting results. Although the discrete cosine transform compression still has a significant effect in the image, by magnifying the image in camera prior to the encode, these images take on a hybrid of pixilation and compression decimation. There is also a slight halation ghosting around the higher contrast edges which get encoded into a pleasant glow.

Read more →
January 28, 2021
Color and the Physiological Response - A Colorist's Exploration
January 28, 2021
Color and the Physiological Response - A Colorist's Exploration
January 28, 2021

Colors, generally speaking, are instinctively associated with emotions. Working as a colorist and having to manipulate color everyday – I feel that most directors and cinematographers intuitively ask for specific colors to enhance specific emotions on special scenes without even realizing it. Here is a little help on how to actually use color in film making to activate – or not – a physical reaction in your audience.

Read more →
January 28, 2021