Advisor: Prof. Chin-Shyurng Fahn

TEL: 02-2733-3141 # 7425

Location: RB307-3

Designer: Yu-Ta Lin

An Intelligent Image Retrieval System by Integrating Multiple Features with Fuzzy Adaptive Resonance Theory Network


@@In the recent years, computer techniques have progressed with each passing day, and it makes the advent of a digitizing epoch. Therefore, hundreds and thousands of pictures and texts can be reserved forever by digitizing. However, how can let users efficiently and effectively retrieve similar images in a large image database to achieve the advanced goal of the video on demand operation is the reason why scientists work hard for image database methods today. Most of the existing image retrieval systems adopt annotations to help the retrieval of similar images. It usually makes a high successful retrieval ratio by human intervening. But it needs the builder of the image database system to spend much more time to construct the relations between stored images, and it also need the cooperation form users. Hence, the techniques of simply using contents to retrieve images in a large image database have been drawn much attention from the people. In this thesis, we propose an intelligence image retrieval system by using a fuzzy ART neural network to integrate multiple features. Our system can retrieve five types of digital images, such as bi-level images, artificial vector gray-level images, complex background gray-level images, artificial scene color images, and natural scene color images. In the system, we separate color information from an input image first. Then we adopt an edge detector to obtain object edges by means of intensity information in the image. From the edge-detected image, we find the most size contour of the object and remove the background to serve as the main part to represent the whole image content. Of this main part, we can extract four kinds of feature data: moment invariants, Fourier descriptors, color bins, and inside contours number that are used as the indices in our image database. For different types of images, the indices are combined with different weights. Through a fuzzy ART neural network to integrate the indices, our system can cluster similar images automatically after they have been processed and stored in the same way into the image database. The experimental results reveal that our approach is feasible and effective.