Date on Master's Thesis/Doctoral Dissertation

5-2024

Document Type

Master's Thesis

Degree Name

M.S.

Department

Computer Engineering and Computer Science

Degree Program

Computer Science, MS

Committee Chair

Frigui, Hichem

Committee Co-Chair (if applicable)

Nasraoui, Olfa

Committee Member

Nasraoui, Olfa

Committee Member

Imam, Ibrahim

Committee Member

Chou, Keven

Author's Keywords

Bag of prototypes; computer vision; automatic target detection; clustering

Abstract

An Automatic Target Detection (ATD) algorithm is capable of identifying the location of targets of interest captured by Infra-Red imagery in vastly different contexts. ATD is often a precursor in a 2-stage methodology in order to ascertain the location and nature of a target in both military and civilian applications. In order to train an ATD algorithm, a large amount of data from varied sources is required. One drawback of this requirement is that some sources of data may harm the performance of the method in different contexts. This thesis explores utilizing an unsupervised method to identify a subset of training data that optimizes the performance of a given test dataset. The methodology summarizes all available data by a set of visual prototypes. These prototypes are then used to map any collection of data into a single vector. The similarity between two collections can be quantified by comparing their mapped feature vectors using Jenson-Shannon (JS) Divergence. We validate this methodology using a large Infra-Red object detection dataset consisting of various training and testing collections. For each test collection, we train the ATD using only the most similar training collections that have the lowest JS divergence. We show that exclusing dissimilar collections from training can improve the trained ATD model.

Share

COinS