Imagegenie: A magic wand for images using deep learning

Sue, Chen Xiang (2023) Imagegenie: A magic wand for images using deep learning. Project Report. Melaka, Malaysia, Universiti Teknikal Malaysia Melaka. (Submitted)

	Text (24 Pages) Imagegenie A magic wand for images using deep learning.pdf - Submitted Version Download (307kB)
	Text (Full Text) Imagegenie A magic wand for images using deep learning.pdf - Submitted Version Restricted to Registered users only Download (3MB)

Abstract

The goal of this project is to develop an application that can increase the resolution of low quality images and to perform neural style transfer that composes one image in the style of another image. In today's digital age, images play a crucial role in communication and self-expression. However, low-quality images are a common problem that many people encounter, particularly when taking pictures with their mobile phones. The lack of image quality can result in images that are blurry, pixelated, or have poor resolution. These low-quality images can be frustrating for users who want to share their images on social media platforms or use them for personal or professional purposes. In this project, several pretrained deep convolutional neural networks model that are the Enhanced Deep Residual Network (EDSR), Efficient Sub-Pixel Convolutional Neural Network (ESPCN) and deep Laplacian Pyramid Super-Resolution Network (LapSRN) are used for increasing the resolution of image and the result of each model is analyzed and compared. The EDSR was chosen as the best model as it achieved the highest average PSNR and SSIM in testing 45 images which are 25.82 dB and 0.70 respectively. Traditional image processing applications often lack the ability to perform advanced image enhancement techniques such as neural style transfer, which can be used to create artistic effects on images. The neural style transfer functionality is achieved by extracting the style of the style image using the VGG19 network architecture which is a pretrained image classification network and apply to the content image to create artistic effects on images. VGG19 was employed because it obtained the highest average ArtFID in 30 testing images which is 45.97 when compare to MobileNet and ResNet. Finally, an application is built using Flutter that combines all the functions above.

Actions (login required)

View Item