Coloring Old images



Taking an old B&W  image and coloring it has become a thing in recent years. When done well it can add life to a image and bring new insights. However generally it takes a lot of work in photoshop and a trained technician. However even done well, it can have issues because often the colorer has to guess on the color of clothes etc

I was interested to learn therefore of work based on machine learning to do the same thing. The idea is to train a algorithm based on a set of images and then use the result to color unrelated images.

I was skeptical initially, but the results looked good so I thought I would have a closer look. I wanted to see if I could run the process myself. The code can be found here

https://github.com/jantic/DeOldify

and instructions are here

https://github.com/jantic/DeOldify/blob/master/README.md

 However it looked like a lot of effort to set it up, but fortunately there was an easier way

If you go to this link

https://colab.research.google.com/github/jantic/DeOldify/blob/master/ImageColorizerColab.ipynb

You can run a virtual machine totally in your browser. If you click it you will see  a page like this

All you need to do is click the boxes in the right hand column. Each time it will do some part of the setup is carried out. Wait till is complete then do the next one

Finally you will reach this section



This is where you do your work. First you need to provide an image.

At this point I have not found a way to load an image from disk so you need to upload your images to the internet. An important fact however is that the URL of the image must end in .jpg. I use imgur.com as my repository which is free. Upload your image and copy the URL into the source url box and then hit the start button under Colorize. After a few minutes it should present you with a screen containing a full resoloution coloured image, a smaller version and the original B&W

I have not found a way of saving the image directly, so instead right click on the full res version and hit save as. You can then save it as a .png file

If you have problems, try reducing the size of the original image.

The results


The results vary depending on the source of the original.  The better the scan, generally the better the image, however the compare well with the hand created version. Plus its free

You can also play with the render factor controls.














Comments

Popular posts from this blog

A Bridgen too far

A new case study

Getting a NVME card ready