Current Trends in Image Quality Perception

Current Trends in Image Quality Perception

Mason Macklem

Simon Fraser University

http://www.cecm.sfu.ca/~msmackle

General Outline

• Examine model of human visual system (HVS)

• Examine properties of human perception of images– consider top-down/bottom-up distinction

• Discuss combinations of current models, based on different perceptual phenomena

Quality-based Model

Quality-based model

• Pros:– Very nice theoretically

– Clearly-defined notions of quality

– Based on theory of cognitive human vision

– Flexible for application-specific model

• Cons:– Practical to

implement?

– Subject-specific definition of quality

– Subjects more accurate at determining relative vs. absolute measurement

Simplified approach

Quality vs. Fidelity

Perception vs Semantic Processing

• Based on properties of HVS

• Models eye’s reaction to various stimuli– eg. mach band, sine

grating, Gabor patch

• Assumes linear model to extend tests to complex images

• Based on properties of Human Attention

• Models subjects’ reactions to different types of image content– eg. Complex, natural

images

• Bypasses responses to artificial stimuli

Human Visual System Model

• Breaks process of image-processing into interaction of contrast information with various parts of the eye

• Motivates representation by discrete filters

• Cornea and lens focus light onto retina

• Retina consists of millions of rods and cones– rods: low-light vision

– cones: normal lighting

– rods:cones => 60:1

• Fovea consists of densely packed cones– processing focusses on

foveal signals

Motivation for Frequency Response Model

• Errors in image reconstruction are differences in pixel values– Interpreted visually as differences in luminance

and contrast values (ie. physical differences)

• Model visual response to luminance and localized contrast to predict visible errors– assuming linear system, measurable using

response to simple phenomena

Visible Differences Predictor (VDP)

Scott Daly

Contrast Sensitivity Function (CSF)

• increasing frequency levels can be resolved to limited extent

• CSF: represents limitations on detecting differences in increasing frequency stimuli– specific to given lens and viewing conditions

• Derive by capturing images for increasing frequency gratings

Common Test Stimuli

Sine grating Gabor patch Mach band

Some Common CSFs

Daly’s CSF (VDP)

Cortex Transform

• Used to simulate sensitivity of visual cortex to orientation and frequency

• Splits frequency domain into 31 (?) sections, each of which is inverse transformed separately

Masking Filter• Nonlinear filter to simulate masking due to local

contrast– function of background contrast

• Masking calculated separately using reactions to sine grating and Gaussian noise

• Uses learning model to simulate prediction of background noise– similar noise across images lessens overall effect

Probability Summation• Describes the increase in the probability of

detection as the signal contrast increases• Calculates contrast difference between the two

images, for each of the (31) images• In most cases, the signs will agree in every pixel

for each cortex band – use the agreed sign as the sign of the probability

• Overall probability is product over all (31) cortex transformed images

• See book for example of Detection Map

Bottom-up vs. Top-down

• Stimulus driven– eg. Search based on

motion, colour, etc.

• Useful for efficient search

• Attracted to objects rather than regions– attention driven by

object properties

• Task/motivation-based– eg. Search based on

interpreting content

• Not as noticeable during search

• Motivation-based search still shows effects of object properties

Saccades & Drifts• Rapid eye movements

– occur 2-3 times/second

• HVS responds to changes in stimuli

• Saccades: search for new ROI, or refocus on current ROI

• Drifts: slow movement away from centre of ROI to refresh image on retina

Veronique

Ruggirello

Influences of Visual Attention

• Measured with visual search experiments– subjects search for target item from group

– target item present in half of samples

• Two measures:– Reaction Time: time to find object correctly vs. number

of objects in set

– Accuracy: frequency of correct response vs. display time of stimulus

• Efficient test: reaction time independent of set size

ContrastEOS increases with increasing contrast relative to background

SizeEOS increases as size difference increases

LocationEOS increases when desired objects are located near center

Even when image content is not centrally located, natural tendency is to focus on center of image

ShapeEOS increases as shape-difference “increases”

Spatial DepthEOS increases as spatial depth increases

Motivation/Context

Where was this photo taken?

Who is this guy?

PeopleAttention more sensitive to human shapes than inanimate objects

ComplexityEOS increases as complexity of background decreases

Other features

• Color: – EOS will increase as color-difference increases– Eg. Levi’s patch on jeans

• Edges:– Edges attended more than textured regions

• Predictability:– Attention directed towards familiar objects

• Motion:– EOS will increase as motion-difference increases

Region-of-Interest Importance Map (ROI)

• Visual attraction directed to objects, rather than regions

• Treats image as a collection of objects– Weights error w/i objects according to various

types of attentive processes

• Results in Importance Map– Weights correspond to probability that location

will be attended directly

ROI Design Model

Image Segmentation

Contrast

Size

Shape

Location

Background/Foreground

W. Osberger

Notes on ROI• VDP Detection Map: probability that existing

pixel differences will be detected• ROI Importance Map: probability that existing

visible pixel differences will be attended• Overall probability of detection should be a

combination of both factors• Open question: single number for either model?