Modeling of human perception for near-threshold local distortion in natural images

Yucheng Liu, Purdue University

Abstract

The ability to predict the perceptual distortion in natural images plays an important role in many modern image/video processing tasks, such as image compression, digital watermarking, and image quality assessment (IQA). Despite the fact that image quality/similarity metric has been investigated intensively over the past few decades, current technique is still far from complete to model the sophisticated interaction between the Human Visual System (HVS) and natural images. Although many metrics have been developed over the years to predict the subjective quality of images with supra-threshold distortions, far less attention is drawn to the perceptibility of distortions at near-threshold level. In previous research, the distortion visibility is known to be related to multiple factors, including but not limited to, light adaptation, contrast sensitivity, and masking effects of various kinds. In this research, we study computational models that predict the visibility of near-threshold distortions in natural images. Our study is conducted in several stages. In the first stage we perform a feature based regression analysis on a local patch basis. For each local fovea-size image patch, we study how various low level statistics of the patch affect the distortion visibility thresholds. An SVM regression is applied to predict the distortion visibility thresholds using the low level features. Feature selection techniques are applied to identify the most relevant features in the threshold prediction. In the second stage, we focus on building an improved perceptual distortion predictor for near-threshold distortions. This involves building a primary cortex (V1) model that predicts low-level responses of the HVS to local distortions at a finer scale than local foveal patch. We consider both conventional gain-control-based method as well as more data-driven methods to build V1 models and evaluate their performances on the public CSIQ masking dataset. We then combine the V1 model and structural familiarity features to generate the final prediction framework, which accounts for the influence of high-level cognition on low-level vision. The proposed framework achieves state-of-art threshold prediction performance on the public CSIQ masking dataset. In the final stage, we apply the complete near-threshold distortion perception model to image compression, whereby we use the perceptual model to guide the spatial distribution of quantization errors across the image to reduce visual impact. Our experiment demonstrates promising results on the evaluation images.

Degree

Ph.D.

Advisors

Allebach, Purdue University.

Subject Area

Electrical engineering|Cognitive psychology

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server
.

Share

COinS