A theoretical comparison of ResNet and DenseNet architectures on the subject of shoreline extraction

Today's Deep Learning technologies provides numerous approaches on the subject of convolutional networks. These approaches serve researchers to train datasets and generate wanted results from these datasets. Each CNN architecture has its own strong points and weak sides. Because of this situation a comparison between these architectures is a valuable asset. Image processing is a method that is frequently used to process remotely sensed data in remote sensing studies.. Between current architectures, RESNET and DENSENET architectures are chosen to be used by Dr. Çavdaroğlu for her project on TÜBİTAK. The result of this comparison will be used in that project in order to apply most efficient architecture. This thesis is written to draw outlines of RESNET and DENSENET and create a foresight for further projects which can be supported by this thesis. In order to achieve an accurate image recognition process in remote sensing domain, a preliminary research is requisite. As a research thesis this work serves the purpose of learning manner of works, performance indicators of RESNET and DENSENET convolutional networks. The result of this research will create a baseline for an academical project. At the other hand, comparison of these two convolutional network approaches provides information to decide which approach is more suitable for remote sensing projects depending upon the subject of the project. For future works on Remote Sensing this thesis work will serve a guideline and reason for preference. The presented thesis work has been developed as the technical feasibility of the 3501 TÜBITAK Project named "Uydu Görüntülerinden Kıyı Sınırlarının Derin Öğrenme Yöntemleri ile Otomatik Çıkarımı", applied by Dr. G. Çiğdem Çavdaroğlu, and the thesis results will be applied within the scope of the project after the project acceptance.