Image Information and Visual Quality Hamid Rahim Sheikh

Image Information and Visual Quality Hamid Rahim Sheikh and Alan C. Bovik IEEE Transactions on Image Processing, Feb. 2006 Presented by Xiaoli Wang for ECE 776 Project

Introduction Quality Assessment (QA) research for image processing “Full-reference” QA methods Interpret image quality as fidelity with a “reference” image Attempt to achieve consistency with the human visual system (HVS) Propose a visual information fidelity measurement for image QA Quantify the loss of image information to the distortion process and explore the relationship between image information and visual quality Natural image source C Channel (Distortion) D HVS E HVS F

Visual Information Fidelity Distortion Model D denotes the random field (RF) from a subband in the reference signal G is a deterministic scalar gain field C stands for the RF from a subband in the reference signal V represents a stationary additive white Gaussian noise RF This model captures important and complementary distortion types: blur, additive noise, and global or local contrast changes Human Visual System Model Approaching the HVS as a “distortion channel” that imposes limits on how much information could flow through it Lumping all sources of HVS uncertainty into an AWGN component

Visual Information Fidelity (cont’) Visual Information Fidelity Criterion (IFC) Mutual information for the reference / test images Represent the information that could ideally be extracted by the brain from a particular subband in the reference and the test images

Visual Information Fidelity (cont’) Visual Information Fidelity Criterion (IFC) (cont’) A simple ratio of the two information measurements relates very well with visual quality Properties of VIF is bounded below by zero VIF is exactly unity if calculated between the reference image and its copy A linear contrast enhancement of the reference image will result in a VIF value larger than unity, signifying a superior visual quality Similarities of VIF with HVS-based methods The numerator is basically IFC (information fidelity criterion) and, hence, is functionally similar to HVS-based methods The denominator can be thought of as a content dependent adjustment

Image samples Reference image, VIF=1. 0 Contrast enhanced, VIF=1. 10 Blurred, VIF=0. 07 JPEG compressed, VIF=0. 10

Experiments DMOS vs. four objective quality criteria Distortion types JPEG 2000 (red) JPEG (green) White noise in RGB space (blue) Gaussian blur (black) Transmission errors in JPEG 2000 stream over fast-fading Rayleigh channel (cyan)

Thanks