Liu Detection

Embed Size (px)

Citation preview

  • 8/3/2019 Liu Detection

    1/19

    A Perceptually Relevant Approach to

    Ringing Region Detection

    Hantao Liu

    Nick Klomp

    Prof. Dr. Ingrid Heynderickx

  • 8/3/2019 Liu Detection

    2/19

    Introduction

    Objective Metrics

    Computational models for the evaluation of image quality

    A Ringing Metric

    To quantify the annoyance of ringing artifacts in compressed images

    Alternatives for expensive image quality assessment by human subjects

    Our main focus is on image and video compression and transmission

    We consider realistic distortions that arise from compression or transmission,

    e.g. blockiness, ringing, blur, and white noise

    We consider the no-reference approach, where the assessment is based onthe compressed image itself only

  • 8/3/2019 Liu Detection

    3/19

    Introduction

    What is Ringing?

    Lossy

    Compression

    Synthetic pattern

    Ringing results from the high frequency component loss due to quantization

    Ringing

    Ringing

  • 8/3/2019 Liu Detection

    4/19

    Introduction

    What is Ringing?

    Compression

    Ringing

    Ringing

    3-D illustration

  • 8/3/2019 Liu Detection

    5/19

    Introduction

    What is Ringing?

    Real-life compressed image

    Ringing

  • 8/3/2019 Liu Detection

    6/19

    Introduction

    Challenges for a Ringing Metric

    Signal dependent noise

    Ringing Detection

    Spatial location of visible ringing artifacts

    Ringing only occurs around high contrast edgesin compressed images

    Spatial maskingof the human visual system (HVS)

    Luminance masking: very dark or very bright areas

    Texture masking: textured or detailed areas

    In agreement with human perception

  • 8/3/2019 Liu Detection

    7/19

    Proposed Algorithm

    Extraction of Perceptual Edges

    Existing methods usually employ an ordinary edge detector (e.g. Sobel)

    An optimal threshold cannot be obtained

    A perceptually more meaningful edge detector is needed

  • 8/3/2019 Liu Detection

    8/19

    Proposed Algorithm

    Extraction of Perceptual Edges

    Smoothing the image until textual details are significantly reduced

    Gaussian Smoothing

    Edge Detector

  • 8/3/2019 Liu Detection

    9/19

    Proposed Algorithm

    Extraction of Perceptual Edges

    Edge-preserving smoothing is needed

    BilateralFiltering

    Line Segment(LS)

    Perceptual Edge Map

  • 8/3/2019 Liu Detection

    10/19

    Proposed Algorithm

    Detection of Perceived Ringing Regions

    Local Region Classification:(1) Edge Region (i.e. EdReg)

    (2) Detection Region (i.e. DeReg)

    (3) Feature Extraction Region (i.e. FeXReg)

    Morphological Operation

  • 8/3/2019 Liu Detection

    11/19

    Proposed Algorithm

    Detection of Perceived Ringing Regions

    Texture masking

    Human Vision Model: Ringing visibility to the human eye

  • 8/3/2019 Liu Detection

    12/19

    Proposed Algorithm

    Detection of Perceived Ringing Regions

    Luminance masking

  • 8/3/2019 Liu Detection

    13/19

    Proposed Algorithm

    Detection of Perceived Ringing Regions

    Implementation of texture masking

  • 8/3/2019 Liu Detection

    14/19

    Proposed Algorithm

    Detection of Perceived Ringing Regions

    Implementation of luminance masking

  • 8/3/2019 Liu Detection

    15/19

    Proposed Algorithm

    Detection of Perceived Ringing Regions

    Computational ringing region (CRR) map

  • 8/3/2019 Liu Detection

    16/19

    Psychovisual Experiment

    Subjective Experiment

    To find out where human beings perceive ringing in natural images

    Visible ringing regions in compressed images were marked by human subjects

    Eight full-color images compressed with JPEG at two compression levels (i.e.16 stimuli) were assessed by 12 subjects (8 males and 4 females)

    Subjective ringing region (SRR) map

    MRR Map (m=12) SRR Map (Thr=1/2)Caps (Q=25)

  • 8/3/2019 Liu Detection

    17/19

    Performance Evaluation

    The computational ringing region (CRR) map predicted by our proposedmethod is compared to the subjective ringing region (SRR) map derived from

    the subjective experiment

    CRR MapSRR Map

  • 8/3/2019 Liu Detection

    18/19

    Performance Evaluation

    The correlation between the CRR and SRR maps is evaluated visually in acomparison map (RGB color image):

    green: correlated ringing regionsred: uncorrelated ringing regions (from CRR)blue: uncorrelated ringing regions (from SRR)

  • 8/3/2019 Liu Detection

    19/19

    Conclusions

    A novelringing region detection method is proposed

    It shows a strong correlationwith subjective data (see [1])

    Current work includes calculating the quantitative correlationbetween theCRR map and SRR map, and comparing our model with alternativesin literature

    Current work includes applying our detection method to a ringing metric

    [1] H. Liu, N. Klomp and I. Heynderickx, "Perceptually Relevant Ringing Region DetectionMethod", EUSIPCO2008 The 16th European Signal Processing Conference, August 2008.

    Reference