Thesis (M.Sc.)-Cairo University, 2023.

Bibliography: pages 54-58.

Image inpainting is filling the missing or corrupted pixels in an image in a realistic
way that cannot be differentiated by human eye. Traditionally, inpainting was done
manually by artists to complete the missing regions in old paintings. In the digital
processing era, image inpainting became an interesting research topic. The used
techniques can be categorized into 2 categories; Non learning techniques and deep
learning-based techniques.
Non-learning techniques have been introduced since the year 2000. These methods
include diffusion-based, patch-based and exemplar-based methods. Diffusion-based
methods use partial differential equations to fill the image holes and ensure the
continuity of edges along it. Patch-based methods search the whole image to find the
perfect patches to complete the image, and Exemplar-based methods tend to merge both
diffusion and patch-based methods.
With the rise of deep learning, it is being widely used in image inpainting. The
used models are capable of studying and learning the structure of the images to
reconstruct the missing regions. Various models are introduced in literature for image
inpainting including simple CNNs, autoencoders, GANs, DCGANs, and multi-stage
networks. These models vary in the size, number of layers and number of parameters in
the model.
Non learning methods require simpler calculations, but they are only suitable for
recovering images with simple structure and small missing regions. Deep learning-
based methods have proven to be efficient for batch processing, and to fill holes of
different sizes with better quality than non learning methods. But deep learning models
require massive processing capabilities and long period of time for training, which may
not be suitable in all cases. In this thesis, we access the complexity issue of training the
image inpainting deep learning models. We propose an autoencoder architecture with
some features added as skip connections, Adam optimizer and leaky ReLU, it has
proven to outperforms other deep learning techniques in literature methods with lower
processing and time complexity.

ترميم الصور هي عملية إكمال الأجزاء الناقصة أو المدمرة من الصور بطريقة واقعية بحيث لا تستطيع العين التمييز بين الأجزاء الأصلية والأجزاء المرممة. التعلم العميق يستخدم بكثرة في ترميم الصور لأن له أداء أفضل من طرق الترميم التقليدية، ولكنه يحتاج إلى موارد معالجة ذات إمكانيات عالية ووقت أطول لتدريب النموذج المستخدم. النموذج المقترح يستخدم المشفر الآلي لترميم الصور، مع بعض التعديلات أثبت هذا النموذج أنه أفضل من بعض النماذج المستخدمة الأخرى من حيث إمكانيات المعالجة والوقت المستهلك.

Issued also as CD

Text in English and abstract in Arabic & English.