Date on Master's Thesis/Doctoral Dissertation

12-2020

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

Amini, Amir

Author's Keywords

Automatic target recognition; convolutional neural networks; spatial transformer networks; wide residual neural networks; DSIAC

Abstract

Automatic Target Recognition (ATR) characterizes the ability for an algorithm or device to identify targets or other objects based on data obtained from sensors, being commonly thermal. ATR is an important technology for both civilian and military computer vision applications. However, the current level of performance that is available is largely deficient compared to the requirements. This is mainly due to the difficulty of acquiring targets in realistic environments, and also to limitations of the distribution of classified data to the academic community for research purposes. This thesis proposes to solve the ATR task using Convolutional Neural Networks (CNN). We present three learning approaches using WideResNet-28-2\cite{wrn} as a backbone CNN. The first method uses random initialization of the network weights. The second method explores transfer learning. Finally, the third approach relies on spatial transformer networks \cite{stn} to enhance the geometric invariance of the model. To validate, analyze and compare our three proposed models, we use a large-scale real benchmark dataset that includes civilian and military vehicles. These targets are captured at different viewing angles, different resolutions, and different times of the day. We evaluate the effectiveness of our methods by studying their robustness to realistic case scenarios where no ground truth data is available and targets are automatically detected. We show that the method that uses spatial transformer networks achieves the best results and demonstrates the most robustness to various perturbations that can be encountered in real applications.

Share

COinS