Abstract

A no-reference image quality assessment (NR-IQA) technique can measure the visual distortion in an image without any reference image data. NR-IQA aims to predict the image quality based on the quality perceived by the Human Visual System (HVS). Image distortions can be caused through the acquisition, compression or transmission of digital images. From the several types of image distortions, JPEG and JPEG2000 compression distortions, addition of white noise, Gaussian blur, and fast fading are the most common. Several approaches were proposed to tackle this problem, some were distortion specific and some were general purpose. Of these, Convolutional Neural Networks (CNN) based approaches have proven to be efficient in predicting quality of the images. Most of these models are trained and tested only for single distortion general purpose images, but in the real world, the images contain more than one distortion type. This Work mainly focusses on using deep convolutional neural networks (DCNNs) for NR-IQA, identifying the different distortion types that are present in the image using distortion type classifiers and also, find the distortion quality of each distortion types using a network of DCNNs. We name this novel approach to be multiple DCNN (MDCNN). We fine tune the networks with different activation functions, optimizers and different tunable parameters in CNNs for the better accuracy. Also, we experiment on different patch sizes that can affect the performance of the system. This proposed model is trained on the LIVE II database and its performance is tested on the CSIQ, and TID 2008 databases which are single distortion. These models achieved high correlation coefficients and accuracy scores on these databases. We further provide the visualization of the inner layers of the DCNN for better understanding of the image quality.

Date of publication

Spring 4-22-2019

Document Type

Thesis

Language

english

Persistent identifier

http://hdl.handle.net/10950/1316

Committee members

Hassan El-Kishky, Mukul V. Shirvaikar, Ron J. Piper, Premananda Indic

Degree

Master of Science in Electrical Engineering

Download
COinS