A study on an improved just noticeable distortion modeling method for images in DCT domain based on frequency-dependent luminance adaptation and texture complexity

Abstract

Although most of the images and videos are consumed through the human visual system (HVS), the current distortion assessment methods adapt an objective methods, such as MSE and PSNR. While these objective methods have an advantage that they can be handled easily in a statistical and mathematical sense but can not reflect HVS characteristics well. Just Noticeable Distortion (JND) means the maximum distortion that human can not perceive. The JND models can be applicable to a variety of image and video processing fields, such as objective quality assessment metrics, video conference, water marking and video compression etc. In this study, we propose a novel spatial JND model that can well reflect HVS characteristics to frequency-dependent luminance adaptation and texture complexity. The proposed JND model has the two main contributions: Firstly, this thesis has obtained more realistic JND thresholds through the various perceptual experiments using real world images depending on frequency, image texture complexity and viewing distances. Through the perceptual experiments, we found out that the conventional JND models have some limitations in representing the perception characteristics of HVS. That is, the conventional contrast sensitivity function (CSF) JND model is not well agreed with the real CSF JND thresholds. Also, we found out that the conventional luminance adaptation effect (LA) JND model yields inaccurate JND thresholds in high frequency regions. Besides, we obtained a more elaborate contrast masking (CM) JND thresholds from our experiments; Secondly, we proposed a combined JND model that takes into account the three factors of CSF, LA and CM. Regarding the CSF JND, we show the limitation of the conventional CSF JND model and propose a new CSF JND model highly suitable to practical JND thresholds. The LA JND and CM JND in the proposed combined JND model are also improved, which better reflects the perception characteristics of HVS for luminance adapta...