A Classification Committee Approach for Improving the Accuracy of Image Query Systems

Publication Year:
2014

No metrics available.

Repository URL:
https://digitalcommons.lmu.edu/etd/335
Author(s):
Almaarik, Abdulaziz
Tags:
Engineering; Systems Engineering
artifact description
Annotating images with text is one of the approaches used to represent semantic meanings of images. Automatic image annotation is becoming increasingly accepted as the first step in keyword-based web-image search applications. Furthermore, the assigning of keywords to images is increasingly being addressed as a classification problem. However, there is no agreement on the best classification approach to use for image classification. Most research in this domain focuses on selecting one machine learning technique and applying it as part of the annotation algorithm. In this project, we start by reviewing five of the most popular classification methods, namely, Support Vector Machines, Multilayer Back-Propagation Neural Networks, Bagging, Nearest Neighbor and Decision Trees. The goal of this study is to find the optimal way to combine the predictions of the different classifiers into one final decision using a committee voting rule based on the predicted accuracy of each classifier. Ensemble methods have been used in the past to improve classification accuracy in many applications, and it is expected to lead to a similar improvement in automatic image annotation as well.