Machine Learning Approach for Content Based Image Retrieval

Introduction

The concept of content-based image retrieval (CBIR) is a very interesting in computer science field that involves retrieving specific images from small to large size image databases based solely on the feature content of the images. The size of the digital image collection is increasing very rapidly due to the advancement in technological devices. These images are stored digitally and transmitted over the Internet at a very high speed. To retrieve the images based on their content effectively and efficiently is essential for further processing of the images. But how to retrieve the images based on their content? There are few image retrieval systems developed commercially as well as academically. Most of the CBIR systems use example image (query image) for retrieving the images from the database.

One of the most common techniques for adding the images into a database is to store images together with some descriptive text or keywords assigned by human operators. These text or keywords are developed based on most prominent feature of an image. For example, for the image of sunset, the most prominent object will be sun or red/orange colour the top of the image. Image retrievals are performed by matching the query texts with the stored descriptive keywords. The images are ranked based on these keywords for similarity. There are several problems with the keyword based matching as this approach is exclusively text based and no visual properties of underlying data are employed. First of all this is very time consuming process, as text descriptions of image contents have to be assigned and keyed in by human operators, also due to enormous volumes of image data, it is also very subjective and incomplete. Human may find different objects or colours or textures in the same image. Retrieval will fail if the user forms the query based on a different set of keywords, or the query refers to the image contents that were not initially described. An ideal system should allow both a keyword and a concept search, in conjunction with a content-based search. The system receives images that are similar to the users’ query and ranks them on the basis of similarity. Many strategies and algorithms have been proposed for similarity based retrieval from the high dimensional index structure. However, there has been little work on query processing based on natural language and fusion of multiple queries. The main task of the image retrieval system to retrieve images based on user’s query. This query may be in the form of keywords, natural language and/or example image. In most of the CBIR systems, the different features such as colour, texture, shape objects are extracted and used for posing a query in the form of example image. The features of the query image and all other images in the database are compared and the distance between them is calculated using similarity measures.