Full-Reference Edge-Based Objective Quality Assessment of Natural and Screen Content Images

• Main Author: Woei-Tan, Loh
• Format: Thesis
• Language: English
• Published: Universiti Malaysia Sarawak (UNIMAS) 2022
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://ir.unimas.my/id/eprint/38691/8/Thesis%20%28Loh%20Woei%20Tan%29_fulltext.pdf; http://ir.unimas.my/id/eprint/38691/

Nowadays, screen content images (SCIs) are gaining popularity other than natural images (NIs). Quality assessment (QA) methods are needed for these two types of images for better quality of experience. In this thesis, two generalized objective QA methods are proposed for NIs and SCIs, i.e. Curvelet based Method (CurM) and Edge Magnitude and Direction Method (EMaD). The modelling of a generalized QA method that works for both types of images is complicated since NIs and SCIs have dissimilar statistical properties. Moreover, some properties of NIs and SCIs are conflicting to one another and this makes the modelling more challenging. The proposed methods assess the perceptual quality of an image based on gradient information. For the CurM, the gradient information is extracted through Curvelet transform. The coefficients from Curvelet transform denote the gradient information in terms of magnitude and direction. Different from the usual practice, CurM considers the gradient direction in 360 degree. On the other hand, EMD filters the images with Prewitt kernel to obtain the edge information and direction. Through the filter results, the image is classified into low and high gradient regions. For the high gradient regions, they are filtered again with bigger kernel size. After extracting the gradient information from the two methods, the gradient information extracted from reference and targeted images are compared to compute a similarity score. This score indicates the quality of the targeted image compared to the reference image. From the performance comparison, it is shown that the proposed methods could assess the perceived quality of NIs and SCIs with high accuracy where CurM and EMaD achieve the weighted average of 0.9063 and 0.9124 respectively in Spearman correlation coefficients for LIVE, SIQAD, and SCID databases.