You are here: Home News Colourising old B&W Photographs

Colourising old B&W Photographs

Using this software Simon Cooper has added colour to many of the village photographs on our website. The results are amazing; photos taken over 100 years ago look so much more recent, with the colours bringing the subjects to life.

See the gallery of UB History images run through the Ai colour process

I sort of stumbled across this project a few weeks ago. Using Artificial Intelligence to recolour archive images.

Photography up until the about 1960's was predominately in Black and White. The world was shades of grey. Then slowly colour started to make its way into popular media, advertising, television, fashion to the point now where monochrome stands out because of its very absence of colour.

Colour is a sensation. We humans are able to perceive a range of electromagnetic wavelengths between 400 nanometers(blues) — 700nm(reds) as this wonderful mix of colours, tones, saturations.

The Upper Broughton Cricket Club

In only the last few years, without realising it, we've started to forget, or at least not be exposed so much to B&W photography. Computers, phones, TV, newspapers, billboards are all in glorious, accurate colour. It's a riotous, joyous thing to behold in the 2020's.

However, there is a fierce debate raging at the moment by historians who rail against a recent trend to use Artificial Intelligence and other tech tools to bring colour and detail to archive images and film. A quick search of the internet will bring up all sorts of links where material from the early days of the medium have been 'improved' and made more accessible to a modern audience. Much of this work is superficially stunning and has taken considerable time and resources to modernise these early historic films and photographs.

Some good examples are:

I'll start by looking at how the above examples of film restoration are achieved.

The techniques used in this field of AI powered film restoration are firstly to increase the frame rate. Typically, early films ran at between 8 to 12 frames per second (fps). That is to say the camera would shoot that number of photographs per second as the film passed through the camera. Often not that accurately, relying on the operator to maintain a steady speed of the cranking handle. When played back these frames give the impression of motion. We are very easily fooled.

The current cinema camera frame rate has been set at 24fps (25 for UK TV and 30 for US TV) for many years but there is a trend now to run at up to 60fps. This gives a much smoother flow of action, less juddering when the camera moves or objects quickly pass in the frame. In order to go from vintage 12fps to 60fps you have to interpolate or 'invent' the intermediate images. The image is also upscaled, made larger and stabilised to remove shake. Some of the above projects have been upscaled to 4K which is the latest hi-definition TV standard. The originals would have been much, much smaller than this. Scratch and hair removal from the original is another vital part of bringing the image up to what we might expect to see today.

Once upscaled, stabilised and smoothed out the final piece of the illusion is to add colour using AI algorithms.

Back in Victorian times and the early part of the 20th century it was not uncommon for photographs, especially portraits to be hand coloured, usually with watercolour paint or dye applied directly to the surface of the monochrome print image. I've used this technique, it's fun and you don't need much colour to produce a convincing look.

The AI model of adding colour to either still or moving images involves training the software model with large amounts of generic sample images and then refining the results so when it sees a face for example it knows what sort of colour faces generally are. (Let's not get into the perplexing detail of what AI does to certain ethnicities. There's lots of documentation out there about the white bias of AI). The AI model starts to know what colour sky, grass etc is. What a party looks like, what a portrait looks like. The picture is built up until there is sufficient information to be able to apply this Artificial Intelligence to images and senes its never seen before. If the results are good it's learned from that model. If not refine the model and try again. This colour can never be historically accurate. There is absolutely no colour information in the originals to work with. It has to be an educated guess.

A leading project for AI colourisation is DeOldify. They offer an off the shelf service via myherritage which I've not used, it requires site registration and is limited to the number of images which can be processed. I chose the model used by the excellent Old Ireland in Colour project. Their model runs a series of programs a remote 'virtual machine' and allows as many images to be processed as you require, with some control over the rendering. DeOldify is also available on GitHub for the true nurd in you.

I wanted to run this AI model really to examine how well the technology was working. I chose some of the Upper Broughton History Project web site images which you can see here. I think you'll agree they are pretty impressive results. Adding colour makes the shots more accessible, we can relate to these people, soldiers, children, they start to look like us here today.

 

 

 

 

 

 

« December 2020 »
December
MoTuWeThFrSaSu
123456
78910111213
14151617181920
21222324252627
28293031