For computer vision tasks using deep learning, should I worry about image size (e.g. 256 x 256) or PPI (pixels per inch)?

I find that PPI is not discussed in the computer vision/deep learning literature.

What's the best PPI for images of size 256 x 256?


PPI is only relevant for physical objects, like representations on screen or paper. It is pixels per inch, and defines how big an image X by Y pixels appears on that object. So if you have a 256x256 pixel image, and you show it at 256 PPI on a screen it will be 1 inch square if you put a ruler up to it.

Many physical devices have natural PPI characteristics. Screens are often 72 PPI, 100 PPI, or about 300 PPI for "Retina" devices. Laser printers start at 300 PPI and higher.

So if you show a 256x256 image on a 100 PPI screen as a direct mapping of one image pixel to one screen pixel, you'll get an image 2.56" across.

If you zoom into the image you can make it bigger on the screen, and it will now cover more than 256 pixels of the screen. The underlying image data will still be an array of 256x256 values

Scanners are another area where PPI is important. If you scan a 1" image at 300 PPI, you get a 300x300 pixel image. Set your scanner to 1200 DPI and you get a 1200x1200 pixel image, that is much finer detail.

In summary, PPI is irrelevant to CV or ML algorithms because they only care about the number of pixels.


Image resolution and PPI do matter in the context of object detection and recognition. Whether you're using a convolutional neural network, traditional computer vision methods, or just looking with your own eyes, if you have an object that only spans 4 by 4 pixels in an image, then you won't be able to detect it reliably by any means unless it is has a unique and distinctive color within the scene. Even then, another object with the same color at the right distance would be mistaken for it because there simply aren't enough pixels to tell the difference.

The concept behind this is called pixel density. Different computer vision tasks, regardless of their proposed solution, require different minimum pixel densities. If we're interested in humans, for example, here's a great example:

enter image description here


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.