Computer Vision: What are state of the art algorithms for efficient image retrieval under light distortion?

In general, it is dependent on the type of object for which you are trying to detect. It is important to consider the object class, because if it is very general (bird, chair, etc.), then distortions may be the least of your worries. In this case, you will require many different views of many different examples (a classic example is the large variation in types of chairs). General frameworks are certainly improving, and are showing promise (such as Felzenszwalb et al.'s deformable parts model http://people.cs.uchicago.edu/~pff/latent/). SIFT (http://en.wikipedia.org/wiki/Scale-invariant_feature_transform) is still one of the best-in-class general purpose descriptors. If your problem is less abstract, and more in the lines of a specific product, template matching may work for you. Book covers are an example of this. Of course, this assumes there is only very slight distortion. Example, with code: http://nashruddin.com/template-matching-in-opencv-with-example.html Finally, if it is more shape-based (handwritten characters, etc.), shape contexts are a great solution. It requires low clutter, but scale, rotation and lighting are rarely a significant issue. Belongie et al. on Shape Contexts: http://www.eecs.berkeley.edu/Research/Projects/CS/vision/shape/belongie-pami02.pdf Note that there have been many variations and improvements for all of these algorithms, so it's worth seeing the new algorithms folks have built on top of these key works.

The problem you described can be solved using image search algorithms based on local features. This class of algorithms can handle huge image databases and provide accurate results. Please check the following source for more detail: An introduction to the image search algorithms based on local features is available at my blog: http://blog.gregorybegelman.com