Browsing by Author "Erten, Hakan"

Now showing 1 - 2 of 2
  • No Thumbnail Available
    Item
    Generatıng Turkısh Lyrıcs Wıth Long Short Term Memory
    (Ankara Üniversitesi Fen Fakültesi, 2020-06-30) Güzel, Mehmet; Erten, Hakan; Bostancı, Erkan; Bilgisayar Mühendisliği; Mühendislik Fakültesi
    Long Short Term Memory (LSTM) has gained a serious achievement on sequential data which have been used generally videos, text and time-series. In this paper, we aim for generating lyrics with newly created “Turkish Lyrics” dataset. By this time, there have been studies for creating Turkish Lyrics with character-level. Unlike previous studies, we propose to Turkish Lyrics generator working with word-level instead on character-level. Also, for employing LSTM, we can’t send the words as string and words must be vectorized. To vectorize, we tried two ways for encoding the words that are used in dataset and compared them. Firstly, we sample for generating one-hot encoding and then, secondly word-embedding way (Word2Vec). Observational results show us that word- level generation with word-embedding way gives more meaningful and realistic lyrics. Actually, there have not been good results enough to be used for a song because of Turkish Grammar. But, this study encourages authors to work on this field and we do believe that this study will initialize research on this area and lead researchers to contribute to this as well.
  • No Thumbnail Available
    Item
    SAR (sentetik açıklıklı radar) görüntülerinde bölütleme
    (Fen Bilimleri Enstitüsü, 2021) Erten, Hakan; Bostancı, Gazi Erkan; Bilgisayar Mühendisliği
    The aim of this thesis are to explain of the processes of creating a new data set with SAR (Synthetic Aperture Radar) images in detail, to demonstrate of how this new dataset is used with deep learning models and to compare of these models used with Mc Nemar's test. Although SAR images can be accessed freely and easily, these are not convenient to be used directly due to the speckle noise, and also there is almost no free available labeled dataset for scientific research. In this study, we propose a novel process that automatedly creates a dataset and removes the speckle noise, labeling images and using the automatedly-created dataset to enhance semantic segmentation task results with state of the art deep neural networks. Used 3 models are evaluated with Mc Nemar's test. As a result, we achieved an overall pixel accuracy (PA) of 92.23% and a mean Intersection over Union (mIoU) of 70.60%. Beside, to show the effectiveness of our noise removal process, we compare the results of models on speckled noise and noise-free versions of our newly-created dataset.