R, Aparna and Channappayya, Sumohana
(2018)
Deep Learning frameworks for Image Quality
Assessment.
Masters thesis, Indian Institute of Technology Hyderabad.
Abstract
Technology is advancing by the arrival of deep learning and it finds huge application in image
processing also. Deep learning itself sufficient to perform over all the statistical methods. As a
research work, I implemented image quality assessment techniques using deep learning. Here I
proposed two full reference image quality assessment algorithms and two no reference image quality
algorithms. Among the two algorithms on each method, one is in a supervised manner and other is
in an unsupervised manner.
First proposed method is the full reference image quality assessment using autoencoder. Existing
literature shows that statistical features of pristine images will get distorted in presence of distortion.
It will be more advantageous if algorithm itself learns the distortion discriminating features. It will
be more complex if the feature length is more. So autoencoder is trained using a large number of
pristine images. An autoencoder will give the best lower dimensional representation of the input.
It is showed that encoded distance features have good distortion discrimination properties. The
proposed algorithm delivers competitive performance over standard databases.
If we are giving both reference and distorted images to the model and the model learning itself
and gives the scores will reduce the load of extracting features and doing post-processing. But model
should be capable one for discriminating the features by itself. Second method which I proposed is
a full reference and no reference image quality assessment using deep convolutional neural networks.
A network is trained in a supervised manner with subjective scores as targets. The algorithm is
performing e�ciently for the distortions that are learned while training the model.
Last proposed method is a classiffication based no reference image quality assessment. Distortion
level in an image may vary from one region to another region. We may not be able to view distortion
in some part but it may be present in other parts. A classiffication model is able to tell whether a
given input patch is of low quality or high quality. It is shown that aggregate of the patch quality
scores is having a high correlation with the subjective scores.
Actions (login required)
 |
View Item |
Altmetric
Altmetric