Image artifacts, resulting from various sources such as vibration, thermal agitation of electrons during image capture, or image transfer across different internet platforms, can significantly degrade image quality. This degradation can impact numerous applications, from weather forecasting based on satellite images to more critical scenarios, such as medical decisions relying on diagnostic images.
Additionally, poor image quality can drive users away from a platform or device. While it is sometimes possible to mitigate these artifacts by addressing the noise source and improving hardware, this approach isn't always viable. Therefore, software-based solutions and mathematical models are developed to address these challenges.
The study highlights the necessity of learning-based techniques for better image quality
In her thesis, MSc Sheyda Ghanbaralizadeh Bahnemiri explored methods for estimating image quality, with a particular focus on noise, one of the most challenging artifacts found in images. Her research focused on developing no-reference image quality assessment (NR-IQA) metrics that do not require a reference image for quality comparison.
"In many cases, the original image is unavailable, or using the entire reference image is inefficient, making no-reference image quality assessment (NR-IQA) an effective solution," she notes.
The absence of a reference image adds complexity, leading to the use of the human visual system as a benchmark for estimating artifacts and assessing image quality. This approach highlights the necessity of learning-based techniques that can mimic the human visual system and make decisions accordingly.
A new learning method improves the metrics
In the field of learning-based image analysis, Convolutional Neural Networks (CNNs) have been the cornerstone of many approaches. In her thesis, Ghanbaralizadeh Bahnemiri leveraged CNN models combined with specialized data preparation techniques to introduce a new learning method for training the model, both for no-reference image quality assessment and noise estimation. She ultimately integrated her proposed methods to improve the performance of image quality assessment metrics.
Ghanbaralizadeh Bahnemiri began her PhD in 2019 as a member of the Computational Imaging Group under the supervision of Professor Karen Eguiazarian. She is now a Machine Learning Specialist at Planmeca, where she applies her expertise in the field of medical imaging.
Public defence on Friday 20 September
The doctoral dissertation of MSc Sheyda Ghanbaralizadeh Bahnemiri in the field of Computing and Electrical Engineering titled Learning-based Image Quality Assessment will be publicly examined at the Faculty of Information Technology and Telecommunication Science at Tampere University at 12:00 on Friday 20.9.2024 at Hervanta Campus, Tietotalo building, auditorium TB 109. The Opponent will be Professor Marco Carli from Roma Tre University. The Custos will be Professor Karen Eguiazarian from Tampere University.
The doctoral dissertation is available online.
The public defence can be followed via remote connection.