Geri Dön

Integration of structured energy-based models with deep neural networks for image restoration

Başlık çevirisi mevcut değil.

  1. Tez No: 721424
  2. Yazar: FAZIL ALTINEL
  3. Danışmanlar: DR. TAKAYUKİ OKATANİ
  4. Tez Türü: Yüksek Lisans
  5. Konular: Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol, Computer Engineering and Computer Science and Control
  6. Anahtar Kelimeler: Belirtilmemiş.
  7. Yıl: 2022
  8. Dil: İngilizce
  9. Üniversite: Tohoku University
  10. Enstitü: Yurtdışı Enstitü
  11. Ana Bilim Dalı: Belirtilmemiş.
  12. Bilim Dalı: Belirtilmemiş.
  13. Sayfa Sayısı: 60

Özet

Özet yok.

Özet (Çeviri)

Image restoration has been an active research topic in computer vision and pattern recognition fields for the past few decades. Image restoration aims to fill in missing regions of images with visually realistic and semantically plausible content. The crucial part of image restoration task is how to model structural relationship between local regions of natural images. Another important point in the task is to use the modelled structural relationship for estimating pixel intensities of the missing regions. Therefore, modelling structural relationship has been a major concern in previous studies of image restoration. Thus, it is often formulated as a structured prediction problem. There have been several studies on image restoration task that uses convolutional neural networks and generative adversarial networks in a feed-forward manner from an input image to an output inpainted image. However, to the best of our knowledge, there is no study investigating energy-based methods that uses deep networks for image restoration task. Predicting structured outputs using energy-based models has been analyzed in previous studies. In general, structured prediction problems are formulated as energy minimization. Moreover, deep neural networks have been used for structured prediction problem due to its generative power. Motivated by the fact that natural images are highly structured, we aim to show that it is possible to complete missing regions in images using the structure with energy-based models and convolutional neural networks. In order to employ structured energy-based-models with deep neural networks for image restoration task, we proposed a structured image restoration method exploiting an energy based model. In order to learn structural relationship between patterns observed in images and missing regions of the images, we employed an energy-based structured prediction method. Energy function is defined by a simple convolutional neural network which takes an image with missing region and an estimate of the true image as inputs. The convolutional neural network has a specific network architecture which includes two paths i having inter-layer connections between paths. The convolutional neural network outputs energy for input image pairs from its final layer. The structural relationship is learned by minimizing energy function. The experimental results on various benchmark datasets show that our proposed method significantly outperforms the state-of-the-art methods which use Generative Adversarial Networks (GANs). We obtained 21.16 dB peak signal to noise ratio (PSNR) on the Olivetti face dataset compared to 18.92 dB provided by the state-of-the-art method. Moreover, we obtained 28.4 dB PSNR on the SVHN dataset and 23.53 dB on the CelebA dataset, compared to 22.3 dB and 21.3 dB, provided by the state-of-the-art methods, respectively.

Benzer Tezler

  1. Mikroşebekelerde ada mod çalışmanın tespiti ve güç kalitesi olaylarının sınıflandırılması için yapay zekâ tabanlı kontrol yöntemlerinin geliştirilmesi

    Development of artificial intelligence based control methods for detection of islanding conditions and classification of power quality events in microgrids

    ALPER YILMAZ

    Doktora

    Türkçe

    Türkçe

    2023

    Elektrik ve Elektronik MühendisliğiBursa Teknik Üniversitesi

    Elektrik-Elektronik Mühendisliği Ana Bilim Dalı

    DOÇ. DR. GÖKAY BAYRAK

  2. Moleküler haberleşme sistemlerinde alıcı kestirim yöntemleri

    Receiver detection methods on molecular communications systems

    ERGİN ASLAN

    Doktora

    Türkçe

    Türkçe

    2023

    Elektrik ve Elektronik Mühendisliğiİstanbul Teknik Üniversitesi

    Elektronik ve Haberleşme Mühendisliği Ana Bilim Dalı

    PROF. DR. MEHMET ERTUĞRUL ÇELEBİ

  3. Audio visual attention for robots from a developmental perspective

    Gelişimsel perspektiften robotlar için görsel ve işitsel diıkkat

    NADA AL AZZAWI

    Yüksek Lisans

    İngilizce

    İngilizce

    2018

    Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik Üniversitesi

    Bilgisayar Mühendisliği Ana Bilim Dalı

    YRD. DOÇ. DR. GÖKHAN İNCE

  4. Kısa dönemli fotovoltaik güç tahmini için geliştirilen yenilikçi bir hibrit modelin analizi ve uygulaması

    Application and analysis of innovative hybrid model for short-term photovoltaic power forecasting

    ALİ RİZA GÜN

    Doktora

    Türkçe

    Türkçe

    2023

    Elektrik ve Elektronik MühendisliğiBilecik Şeyh Edebali Üniversitesi

    Elektronik ve Bilgisayar Mühendisliği Ana Bilim Dalı

    PROF. DR. MEHMET KURBAN

    DOÇ. DR. EMRAH DOKUR

  5. Directional regularization based variational models for image recovery

    Görüntü onarımı için yön güdümlü düzenlemeye dayalı varyasyonel modeller

    EZGİ DEMİRCAN TÜREYEN

    Doktora

    İngilizce

    İngilizce

    2022

    Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik Üniversitesi

    Bilgisayar Mühendisliği Ana Bilim Dalı

    PROF. DR. MUSTAFA ERSEL KAMAŞAK