Current MS Students/Kelly Breed
From CSWiki
The main goal of query by content (a.k.a. content based image matching) is to retrieve same and/or similar images to a given query image based not on text keywords or manually pre-specified categories but on automatic matching of image features describing image contents in various levels.
My thesis proposes a query by content system to categorize images based on hybrid feature and a self-organizing map (SOM). In order to implement such a system, I will start with an application already built through CS491 and will extend the system in three directions: 1) defining new feature vectors, 2) defining a new similarity measure, and 3) attaching a query system.
The current system organizes images using a SOM (A SOM can be considered as a set of nodes and each node represents a group of similar images.) First of all, the system extracts from each image its features defined by intensity, indexed color, and/or primitive texture information, and it uses these extracted features to find similar images. More precisely, the Euclidean distance between features of two images is used to determine a similarity. At the end, similar images are grouped into a node of the SOM.
By nature, the performance of the current system is significantly influenced by image feature extraction and the definition of similarity measure. For my thesis, I plan to explore a mixture of image features such as color components in different color spaces and wavelets features. These hybrid features will capture color distributions, spatial correlation, and geometric properties and will be used to acquire perceptual properties of an image. Additionally, I plan to look into several similarity measures such as geodesic distance or cross spectrum correlation in order to cope with the limitation of Euclidean distance.
As the last effort, I plan to implement the query system as the current system does not provide the query part. Besides accepting an image as a query, I aim to provide the ability for a user to draw directly in the UI to query for images.
At the end of my studies I will demonstrate a system that analyzes a large number of images (probably hundreds to thousands), organizes them according to extracted features and offers the user a method for finding/searching for an image based on a query image or a hand drawn graphic. A method for determining how accurate the matches that are produced by the application will be reported as well.
