Download Human Visual Perception - Lecture Slides | ECE 6258 and more Study notes Digital Signal Processing in PDF only on Docsity!
9/17/2003 ECE 6258 Russell M. Mersereau 1
ECE6258 Lecture 13
Human Visual Perception
9/17/2003 ECE 6258 Russell M. Mersereau 2
The Compression Problem
n The bitrates achieveable using lossless compression are way too high for most commercial applications. n Lossy compression means that the reproduced images will contain errors. q For a given bit rate, how can we reduce the perceived distortion (improve quality)? q For a given level of perceived distortion, how can we reduce the bit rate? q How big is the gap between “lossless” and “perceptually lossless”?
9/17/2003 ECE 6258 Russell M. Mersereau 3
A General Coding Structure
Visual Source
Human Receiver
Convert to Representation
Invert Representation
Quantize
Dequantize
Entropy Coder
Entropy Decoder
Channel
OVERALL ENCODER
OVERALL DECODER
9/17/2003 ECE 6258 Russell M. Mersereau 4
Issues
- Images do not satisfy traditional statistical assumptions, such as stationarity.
- Tractable distortion measures (mean squared error, SNR, etc.) do not reflect viewer’s subjective quality judgements well.
- Classical compression methods, such as DPCM and transform coding, achieve compression by reducing statistical redundancy in the source.
- Understanding the receiver aids compression by reducing information irrelevancy —there is no need to enhance or compress what cannot be seen!
9/17/2003 ECE 6258 Russell M. Mersereau 5
The Mission
n What does classical psychophysics tell us
about how we perceive (B/W) images?
n How can we use this information to hide
compression errors (or watermarks, secret
messages, etc.)?
9/17/2003 ECE 6258 Russell M. Mersereau 6
n Luminance measures physical light intensity.
q i is the illumination. (0< i ) q r is the reflectivity of the object. (0< r <1) q l is the luminance efficiency function of the recording system.
Luminance
9/17/2003 ECE 6258 Russell M. Mersereau 7
Solar Illumination
n Spectral content of solar radiation at noon in Washington, DC.
q Solid – Above earth’s atmosphere q Dashed – At ground level
(from Hardy, Handbook of Calorimetry , 1936).
9/17/2003 ECE 6258 Russell M. Mersereau 8
Luminance Efficiency
n A typical luminance efficiency function
from Jain, Fundamentals of Digital Image Processing , 1989
9/17/2003 ECE 6258 Russell M. Mersereau 13
Luminance and Brightness
n Luminance measures physical light intensity.
n The perceived intensity, or brightness , depends upon the luminance of both the object and the background. The illusion of simultaneous contrast
The luminance of both center objects is the same
9/17/2003 ECE 6258 Russell M. Mersereau 14
Mach Band Effect
Are the bands of uniform color? Are they uniformly spaced?
9/17/2003 ECE 6258 Russell M. Mersereau 15
Mach Band Effect
n The HVS does not perceive a staircase of uniform luminance steps to consist of uniform brightness steps.
n Can be used to estimate the HVS impulse response.
9/17/2003 ECE 6258 Russell M. Mersereau 16
HVS Impulse Response
n The negative lobes indicate lateral inhibition —one cell can weaken signals from its neighbors. q Suggests that edges can mask noise
n Implications for coding: 1. Sharp edges can degrade slightly. 2. Noise can be hidden reliably within 10 arcmin of the dark side of an edge.
9/17/2003 ECE 6258 Russell M. Mersereau 17
Edge Masking (Facilitation)
Facilitation here
Elevated thresholds here
9/17/2003 ECE 6258 Russell M. Mersereau 18
Contrast Sensitivity Function (CSF)
n Describes the HVS response to sinusoidal stimuli.
n The value of C at which z becomes just noticeable gives the VT as a function of ( f r, θ ).
n The reciprocal of the VT produces the CSF.
9/17/2003 ECE 6258 Russell M. Mersereau 19
Contrast Sensitivity Function
n Bandpass shape (peak at 2—10 cpd). n Max at θ = 0,90o 3 dB down at θ =45o. n Highest frequency detectable 50— cpd. n Lowest frequency detectable 0.5 cpd.
n Implications for coding: 1. Noise can be hidden at high spatial frequencies.
- An images frequency components can be weighted in proportion to HVS sensitivity.
9/17/2003 ECE 6258 Russell M. Mersereau 20
Contrast Masking
n If a mask raises the VT of a target, contrast masking occurs.
n The mask and the target must have similar locations, frequencies, and orientations.
9/17/2003 ECE 6258 Russell M. Mersereau 25
Summary of HVS Observations (1)
n Information Representation
q HVS channels suggest that signal representations with multiscale, multiresolution behavior in the frequency domain (e.g. subband and wavelet decompositions) might be desirable.
q Purely spatial techniques (DPCM, block-based systems, VQ) become less attractive.
9/17/2003 ECE 6258 Russell M. Mersereau 26
Summary of HVS Observations (2)
n New Distortion Measures
q Distortion measures should include local HVS sensitivity. q Errors below the VT should not contribute to distortion. q Errors in different representation components may be weighted according to HVS sensitivity and then used to compute a weighted MDS. q Alternatively, the input data may be mapped via an HVS model into a percuptually flat space when regular MSE can be employed.
9/17/2003 ECE 6258 Russell M. Mersereau 27
Summary of HVS Orservations (3)
n Mask Profiling
q The HVS can be used to estimate the amount of error that is tolerable in each representation component. q Provided that all quantization errors remain below this JND, perceptual losslessness is preserved. n Such a scheme forms a variable-bit rate, constant- quality system. n Rate-control can be used to produce a constant-bit rate, variable-quality system.
