Image similarity search

Abstract

With the increase of digital photography and the increase of image sharing over the internet, an essential means to search through all these images should be provided, to help the user find the required images. Most of the image search engines found on the web use keyword based queries to search for images, thus using a keyword based indexing technique. These keywords should be provided or generated in a way, since images are not made of words. The terms, used to describe the images, are difficult to be automatically generated and hence humans should manually input these terms. Normally the image name is used but this depends on how well the image owner named the image. To solve this problem, our proposed system will use a different approach, than that used in keyword based image search engines. Since an image is not made of terms but made of pixels, the pixels themselves will be used to describe the image. The Image Similarity Search engine will use a region based image retrieval approach, where the image is split into regions and by using the pixels in each region, a region description will be generated. By using these region descriptions an image description can be generated. The regions should represent an object in the image, for example a dog, the sky, the road, etc ... The problem in this approach is that the user can't input these features as a query so instead of using a term based query the system will use another image as the search query. Similar images with similar features will be fetched by the system and shown to the user. This document will describe the design and implementation of this system and the idea behind. An evaluation of the results of the system was also made to see the efficiency of the system.