machine-learningartificial-intelligencecomputer-visionimage-recognitioncbir

Can anyone suggest good algorithms for CBIR?

Project: Content Based Image Retrieval - Semi-supervised (manual tagging is done on images while training)

Description

I have 1000000 images in the database. The training is manual (supervised) - title and tags are provided for each image. Example: coke.jpg Title : Coke Tags : Coke, Can

Using the images and tags, I have to train the system. After training, when I give a new image (already in database/ completely new) the system should output the possible tags the image may belong to and display few images belonging to each tag. The system may also say no match found.

Questions:

1) What is mean by image fingerprint? What is the image fingerprint size expected ? (important because there will be millions of images to be inserted in database)

2) What is the field format of that fingerprint in the database ? (important because a fast search is needed … script should search in a 1M images database in less than 1 second)

3) What is the descriptors (algorithms) we use to analyze them ?

Thanks in advance

Solution

  • Well, this topic is very large, but here is a brief overview of a possible solution

    1. Image fingerprints are collections of SIFT descriptors These are quantized both to reduce size, and to allow indexing

    2. Build an inverted index of your database to allow looking up an image by quantized descriptors (you can use any full text search engine \ DB for this)

    3. Given an image, lookup images which share a large amount of common descriptors

    4. For those potential candidates, you should validate that the spatial arrangement of descriptors is similar enough

    Some articles to get you started:

    Philbin, James, et al. "Object retrieval with large vocabularies and fast spatial matching." Computer Vision and Pattern Recognition, 2007. CVPR'07. IEEE Conference on. IEEE, 2007.

    Philbin, James, et al. "Lost in quantization: Improving particular object retrieval in large scale image databases." Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on. IEEE, 2008.

    Mikulík, Andrej, et al. "Learning a fine vocabulary." Computer Vision–ECCV 2010 (2010): 1-14.