Bilinear CNN for Non-reference Image Quality Prediction

Image quality assessment (IQA) is widely applied in many areas such as camera technology, digital television, stream video service. With the increase of computer networks transmission bit rate, the greater requirements for non-references image quality assessment in practice prompts the researchers to find an accurate algorithm. Image recognition based on convolution neural network (CNN) has proved its superiority in computer vision and digital signal processing tasks. However, applying deep CNN in non-reference image quality assessment (NR-IQA) remains challenges due to the lack of sample database, superabundant sample categories, and other problems.In this thesis, first, we provide an overview of several successful approaches to NR-IQA such as BRESQUE, FRIQUEE, and DIQA. And we obtain the conclusion and comparison of those algorithms’ advantages and disadvantages. In addition, we introduce the Bilinear CNN structures which are well used in image recognition area. The approaches and flow chart of Bilinear CNN are given in detail.Furthermore, we applied the structure of Bilinear CNN to patch-wise and image-wise methods in non-reference image quality prediction. The detailed training progress and CNN networks specifications are given. The performance of prediction and challenge problems in NR-IQA are discussed.

Author

• Zhang, Kehan

Language

• en

Keyword

• Non-reference
• Patch-wise
• Bilinear CNN
• Image quality prediction
• Image-wise

Date created

• 2019

Type of Work

• Master's Thesis

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Electrical Engineering

Degree

• M.S.

Advisor

• Doroslovački, Miloš

Committee Member(s)

• Eom, Kie-Bum
• Helgert, Hermann J.

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/m613mz29d