0

I'm looking for a bit of guidance... I have a single image of a passport photo. What I would like to do is separate the person and the background. I will then analyse the background colour, pattern, etc.

I've done some research into segmentation and most of what I'm finding uses multiple frames. Unfortunately I only have one frame and do not have the luxury to take a picture of the background without a person.

I have also looked at GrabCut, but I do not want the user to have to manually select background/foreground parts of the image. Is there an automated version of this, possibly?

I was thinking about using contour/edge (canny?) detection?...

Thank you in advance!

Edit: Sample set of images can be found here: https://i.stack.imgur.com/FvRgF.jpg

So I followed Froyo's advice and so far have the following result: https://i.stack.imgur.com/yQpdx.png

Some of the edges are cut off because of the face detection rectangle, so I'll need to increase the size of the rect when a face is detected.

LKB
  • 1,020
  • 5
  • 22
  • 46
  • what is your requirement from the image...r u looking for the whole person without background or just face/hand/nose etc .. ?? – Rahul galgali Sep 02 '14 at 07:22
  • There are two main possibilities I see: 1. let the computer learn what a person is. Detect the person and consider everything else as background. This task isn't completely solved yet I guess. 2. make assumptions about what foreground/background is. Main contour (centered and/or biggest contour) of the image is foreground, rest is background. This task might be very erroneous. Maybe you can use ideas of version 2 combined with some learning mechanism? ;) – Micka Sep 02 '14 at 07:48
  • can you provide a set of sample images maybe? – Micka Sep 02 '14 at 07:49
  • What kind of background are you dealing with. Normally pp-sized pics have white or single-color without any noise. If this is the case, It is going to be a lot easier :) – Pervez Alam Sep 02 '14 at 07:56
  • @Rahulgalgali - Thanks for your reply. I am looking for the whole person. I am already conducting feature extraction, just looking to separate the entire body/head of person from image now. – LKB Sep 02 '14 at 07:56
  • @Micka - Thanks for your reply! That makes sense... would definitely need some assumptions about the image, not sure how reliable that would be. I'll edit my main post for a sample set of images now. – LKB Sep 02 '14 at 08:01
  • @PervezAlam - It is designed for passport images currently (the aim is to detect if the background is purely white so meets passport photo standards), however I'm going to accept any type of image further down the track. – LKB Sep 02 '14 at 08:03

1 Answers1

4

You could use GrabCut and wouldn't need to get manual input. You should detect the face using Haar Cascades and pass the bounding box to GrabCut. I think this would work.

I have some experience in segmenting foreground and background, but I also had a disparity map, so it was sort of easier. I used superpixels and clustered them to get separate foreground and background extracted.

SLIC and SLICO are the state of the art algorithms for superpixels segmentation. Here is some of the work done in this field.

https://cvhci.anthropomatik.kit.edu/~baeuml/downloads/Schick2012.pdf

http://www.cs.huji.ac.il/~daphna/papers/Rosenfeld_ICCV2011.pdf

You could also try Saliency Maps. They are used to extract salient objects from the image and in your case the foreground that you are looking for.

Froyo
  • 17,947
  • 8
  • 45
  • 73