Araştırma Çıktıları | WoS | Scopus | TR-Dizin | PubMed

Permanent URI for this communityhttps://hdl.handle.net/20.500.14719/1741

Browse

Start date: 2010End date: 2019Type: search.filters.itemtype.Conference PaperSubject: search.filters.subject.Signal processingSubject: search.filters.subject.Classification (of information)Item Type: PublicationAuthor: search.filters.author.Arica, Nafiz

Search Results

Now showing 1 - 3 of 3
  • No Thumbnail Available
    PublicationMetadata only
    Facial emotion classification using deep embedding with triplet loss function, Üçlü Yitim Fonksiyonu Tabanli Derin Gömme ile YÜz Duygu Siniflandirmasi
    (Institute of Electrical and Electronics Engineers Inc., 2017) Bircanoglu, Cenk; Arica, Nafiz; Bircanoglu, Cenk, Bilgisayar Mühendisliǧi Bölümü, Bahçeşehir Üniversitesi, Istanbul, Turkey; Arica, Nafiz, Bilgisayar Mühendisliǧi Bölümü, Bahçeşehir Üniversitesi, Istanbul, Turkey
    In this paper, a deep embedding method using triplet loss function is proposed for classification of the emotions in face images. The originality of proposed method lies in the loss function, different from the other deep learning-based facial emotion classification approaches. The input face images are embedded into a lower dimensional feature space using a multi-layer convolutional neural network. In this embedding process the loss function is calculated by taking triple samples in the training dataset. For each sample in batches, two samples are selected in such a way that one of them is from the same class and the other one from the different class. The loss function aims to close the samples belonging to the same class to each other in the Euclidean space and to move away from the samples in the different class. In the performance analysis of proposed method, two popular deep architectures, namely AlexNet and VGG are used on two datasets called CIFE and GaMo. The experiments show that the proposed method outperforms the other popular loss functions and the available studies on this particular datasets. © 2017 Elsevier B.V., All rights reserved.
  • No Thumbnail Available
    PublicationMetadata only
    A case study on transfer learning in convolutional neural networks, Evrişimli sinir aǧlarinda eǧitim transferi için örnek çalişma
    (Institute of Electrical and Electronics Engineers Inc., 2018) Gürkaynak, Cahit Deniz; Arica, Nafiz; Gürkaynak, Cahit Deniz, Bahçeşehir Üniversitesi, Istanbul, Turkey; Arica, Nafiz, Bahçeşehir Üniversitesi, Istanbul, Turkey
    In this work, a case study is performed on transfer learning approach in convolutional neural networks. Transfer learning parameters are examined on AlexNet, VGGNet and ResNet architectures for marine vessel classification task on MARVEL dataset. The results confirmed that transferring the parameter values of the first layers and fine-tuning the other layers, whose weights are initialized from pre-trained weights, performs better than training network from scratch. It's also observed that preprocessing and regularization improves overall scores significantly. © 2018 Elsevier B.V., All rights reserved.
  • No Thumbnail Available
    PublicationMetadata only
    A comparison of activation functions in artificial neural networks, Yapay sinir aǧlarinda aktivasyon fonksiyonlarinin karşilaştirilmasi
    (Institute of Electrical and Electronics Engineers Inc., 2018) Bircanoglu, Cenk; Arica, Nafiz; Bircanoglu, Cenk, Bilgisayar Mühendisliǧi Bölümü, Bahçeşehir Üniversitesi, Istanbul, Turkey; Arica, Nafiz, Bilgisayar Mühendisliǧi Bölümü, Bahçeşehir Üniversitesi, Istanbul, Turkey
    In this study, the effects of Activation Functions (AF) in Artificial Neural Network (ANN) on regression and classification performance are compared. In comparisons, success rates in test data and duration of training are evaluated for both problems. A total of 11 AF functions, 10 AF commonly used in the literature and Square function proposed in this study, are compared using 7 different datasets, 2 for regression and 5 for classification. 3 different ANN architectures, which are considered to be the most appropriate for each dataset are employed in the experiments. As a result of totally 231 different training procedures, the effects of Afs are examined for different datasets and architectures. Similarly, the effects of AF on training time are shown for different datasets. In the experiments it is shown that ReLU is the most succesfull AF in general purposes. In addition to ReLU, Square function gives the better results in image datasets. © 2018 Elsevier B.V., All rights reserved.