Deep convolutional neural networks with an application towards geospatial object recognition

---

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Mühendislik Fakültesi, Elektrik ve Elektronik Mühendisliği Bölümü, Türkiye

Tezin Onay Tarihi: 2014

Öğrenci: EMRECAN BATI

Danışman: ABDULLAH AYDIN ALATAN

Özet:

The passion of human-being to invent intelligent systems becomes more and more meaningful day by day, as the data captured every second by artificial sensors needs to be examined and classified for many applications. The processing of ever-increasing amount of data by defining information explicitly seems nearly impossible, regarding the variability and the amount of the information, which reveals the need for intelligent systems that are capable of learning. Deep learning is a set of algorithms that attempts to find a hierarchical representation of the input data by trying to mimic the way human brain captures the critical aspects of excessive sensory data, to which it is exposed to every second. Convolutional neural networks, which are trainable learning structures, are also biologically inspired from the receptive fields in visual cortex. In this thesis, the performance of convolutional neural networks are investigated for an application towards geospatial target detection and classification from satellite images. Based on the experiments, it is observed that the utilization of preprocessing, dropout, i.e. dropping neurons randomly in the training phase, and rectified linear unit as the activation function improves the classification rate, significantly. However, the application of this deep classifier on satellite images still yields high false alarm rate, possibly due to insufficient number of training data.