[attachment=15220]
[attachment=15221]
[attachment=15222]
[attachment=15223]
1. INTRODUCTION
There are number of existing retrieval techniques for images. Text based techniques are the most
commonly used. In text-based systems the images are first annotated by text and then text indexing and
retrieval techniques are used. Content based retrieval techniques followed the text-based system where
in low level image features like color, texture etc. are used to retrieve similar images. In text-based
systems, manual annotation poses a difficulty and in visual-based systems, subjectivity of human
perception makes it a tough task. The low level features fail to capture the high level semantics of the
image in content based systems. To help capture the high level semantics, human is looped in and the
process of relevance feedback is used to make the system adapt to the user.
Retrieval techniques used for search try to find the most similar images given to user input in the form
of text, visual or other features. But in case the aim of user is to explore the set of images, s/he might
not know the description of the image. Returning the most similar images as results of the initial query
might not be best idea. Catering to the needs of users who want to explore rather than search is one of
the human centered research areas in multimedia.
In this project, a Content-based Image Retrieval system for finding similar images in personalized photo
collections on local machine is described. Instead of matching low level features, the attempt is to
match the objects in various images. The image is treated as the sum of objects of interest and
remaining background. User can register important regions in images which in essence are the objects of
interest. In a personalized collection, the objects of interest can be broadly classified into - people and
places. The SIFT features have been found to be useful in recognizing faces and is by definition scale and
rotation invariant hence making it suitable to detect rigid objects in the background. Hence for every
image in the database, the best match for all registered regions is stored and the MPEG 7 color and
texture features are stored for the remaining background. MPEG 7 features have been extensively used
in practice for content based multimedia retrieval. The flowchart of the system, description of the
components and the details of the algorithms to be used are discussed in this report.
2. MOTIVATION
In a diversity of areas ranging from medical, astronomy, geology, military and so on, the applications of
imaging techniques have been ever increasing. Image acquisition, storing, transfer and retrieval
techniques have undergone tremendous improvement over the past few decades. The common
property underlying the aim of evaluating images irrespective of the field of interest is the need to
explore the image and infer valuable information from it. Thus, the semantics of the image are far more
important than the features of image which brings us to biggest challenge in image retrieval and
exploration applications – bridging the semantic gap. It refers to the intelligent use of low level features
and well defined metadata to understand the semantics of the image. Since an image usually intends to
capture an object in particular, for example, brain scans in MR imaging, satellite imagery of area under
military surveillance or the simply the personal photographs of any person or place.