Date on Master's Thesis/Doctoral Dissertation


Document Type

Master's Thesis

Degree Name



Computer Engineering and Computer Science

Degree Program

Computer Science, MS

Committee Chair

Frigui, Hichem

Committee Member

Nasraoui, Olfa

Committee Member

Xiang, Zhang

Author's Keywords

Automatic; target; recognition; deep; metric; learning


An Automatic Target Recognizer (ATR) is a real or near-real time understanding system where its input (images, signals) are obtained from sensors and its output is the detected and recognized target. ATR is an important task in many civilian and military computer vision applications. The used sensors, such as infrared (IR) imagery, enlarge our knowledge of the surrounding environment, especially at night as they provide continuous surveillance. However, ATR based on IR faces major challenges such as meteorological conditions, scale and viewpoint invariance. In this thesis, we propose solutions that are based on Deep Metric Learning (DML). DML is a technique that has been recently proposed to learn a transformation to a representation space (embedding space) in end-to-end manner based on convolutional neural networks. We explore three distinct approaches. The first one, is based on optimizing a loss function based on a set of triplets [47]. The second one is based on a method that aims to capture the explicit distributions of the different classes in the transformation space [45]. The third method aims to learn a compact hyper-spherical embedding based on Von Mises-Fisher distribution [64]. For these methods, we propose strategies to select and update the constraints to reduce the intra-class variations and increase the inter-class variations. To validate, analyze and compare the three different DML approaches, we use a large real benchmark data that contain multiple target classes of military and civilian vehicles. These targets are captured at different viewing angles, different ranges, and different times of the day. We validate the effectiveness of these methods by evaluating their classification performance as well as analyzing the compactness of their learned features. We show that the three considered methods can learn models that achieve their objectives.