Self-supervised building detection with decision fusion

---

Tezin Türü: Doktora

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Mühendislik Fakültesi, Bilgisayar Mühendisliği Bölümü, Türkiye

Tezin Onay Tarihi: 2013

Öğrenci: ÇAĞLAR ŞENARAS

Eş Danışman: PEKİN ERHAN EREN, FATOŞ TUNAY YARMAN VURAL

Özet:

This thesis proposes a new building detection framework for monocular satellite images, called Self-Supervised Decision Fusion (SSDF). The model is based on the idea of self-supervision, which aims to generate training data automatically from each individual test image, without any human interaction. This principle allows us to use the advantages of the supervised classifiers in a fully automated framework. The technical shortcomings of the available supervised and unsupervised algorithms, such as difficulties in manual labeling of the images to extract the training data, large interclass variances and a wide variety of buildings, prevent the previous studies to satisfy the need of robust autonomous detection systems. We attempt to overcome these problems by combining our previous supervised and unsupervised building detection frameworks to suggest a self-supervised learning architecture. We borrow the major strength of the unsupervised approaches in order to obtain one of the most important clues, the relation of a building and its cast shadow in order to solve the major problem of training of the supervised approaches. Furthermore, supervised study allows us to combine the detection results of multiple classifiers under a hierarchical architecture, called Fuzzy Stacked Generalization (FSG). The suggested method involves three major steps: In the first step, after pan-sharpening and segmentation process several masks are extracted to represent the invariant information about the building object. These masks are vegetation, shadow and rectangular structure masks. In the second step, by employing these masks negative and positive samples are selected from each image layout. Finally, the training data extracted in the second step is used to train FSG.