Hide/Show Apps

Visibility of dct quantization noise for high dynamic range images.

Sözer, Sevim Begü
Discrete cosine transform (DCT) is one of the fundamental stages, which is widely used and studied in lossy image compression. This technique aims to generate a minimum visible distortion in an image with respect to human visual system (HVS). The studies until now propose DCT quantization matrices for low dynamic range (LDR) images to display on LDR monitors. In this thesis research, the visibility of DCT quantization noise is studied for high dynamic range (HDR) images. High luminance patterns for DCT quantization noise are formed with mean luminance levels of 100 cd/m2, 500 cd/m2, 1000 cd/m2 and 1500 cd/m2, for various DCT frequencies and amplitude levels. A novel rendering algorithm is proposed to generate the high luminance patterns for DCT quantization noise by using an HDR monitor (SIM2 HDR47). The visibility of DCT quantization noise is measured with psycho-visual experiments based on geometric search and QUEST methodologies. The estimated quantization levels with geometric search is utilized to form the initial statistics for the QUEST search. The quantization levels for each frequency are then obtained by fitting the QUEST results to a Weibull function. The experimental results first indicate that HVS is more sensitive to high DCT frequencies and high mean luminance levels. Second, although increased luminance levels cause increased quantization levels, they cause a decrease in contrast threshold values. Finally, both the geometric search and QUEST methodologies are observed to produce coherent results in the experiments.