hue-grayscale collaborating edge detection & edge color distribution space jiqiang song march 6...

Hue-Grayscale Collaborating Edge Detection & Edge Color Distribution Space

Jiqiang Song

March 6th, 2002

Introduction

• Definition of “Edge” in an image– Shape transition of intensity and/or color

• Meaning of edge– Outline of objects– Image structure– An important feature for image segmentation &

object detection

Part 1: Hue-Edge Collaborating (HGC) edge detection

• Existing edge detection methods– Binary image

– Grayscale image

– Color image

Binary edge detector

• A foreground pixel ‘P’ (P=1) is an edge point if its convolution result does not equal zero.

HGC Edge Detector —

P2

P1

P3

P8

P

P4

P7

P5

P6

0

-1

-1

0

4

0

-1

-1

0

• 0

P2

P1

P3

P8

P

P4

P7

P5

P6

-1

-1

-1

-1

8

-1

-1

-1

-1

• 0

8-connected edges

4-connected edges

Grayscale edge detector

• Gradient operators– Sobel, Prewitt, Roberts

• Second derivative operators– Zero-crossing, LoG

• Others– Canny, SUSAN

HGC Edge Detector —

Color edge detector• Multi-dimensional gradient methods

• Output fusion methods

HGC Edge Detector —

R

G

B

Multi-dimensional gradient calculation

Thresholding Color edges

R

G

B

1D Edge detection

Output fusion Color edges

1D Edge detection

1D Edge detection

Why to design a HGC edge detector?

• Grayscale edge detector >90% of real

edges, fast.

• Color edge detector more edges, slow.

• Our application: video processing

– Thousands of images in a 10 minutes long

video (when sampling 3~4 images/second)

– Color edge detector often over-detects edges.

HGC Edge Detector —

Introduction of color models

• RGB– R (red); G (green); B (blue)

• Grayscale– Luminance, achromatic, 1 dimension

• HSI – a perceptual color model– H (hue); S (saturation); I (intensity)

• Others: YUV, HIQ, CIE(Lab),…

HGC Edge Detector —

Grayscale vs. HSI

• RGB Grayscaleg = 0.299R + 0.587G + 0.114B; (0 g 1)

• RGB HSI

HGC Edge Detector —

Grayscale vs. HSI (continued)

1. The change of hue cannot be detected in grayscale space.

2. The noticeable change of intensity or saturation can be detected in grayscale space.

HGC Edge Detector —

HGC edge detector

Step 1: Generate Hue Edge Map (HEM) & Grayscale Edge Map (GEM)

Step 2: Overdetected edge minimization

Step 3: Output fusion

HGC Edge Detector —

Hue Edge Map & Grayscale Edge map• Convert a sampled RGB video image into a hue map &

a grayscale map.• Use Sobel operator to detect edge strength (gradient) in

two maps.• Use a fuzzy threshold to generate edge maps.

HGC Edge Detector —

Overdetected hue edge minimizationASSUME: a valuable edge point must have a certain

connected length.• Extract hue edge points that are not grayscale edge

points.

• Use a run-length transform (RLT) to calculate the maximum connected length of an edge point in any direction.

• Remove edge points that are not of desired connected length.

HGC Edge Detector —

Output fusion

• Merge HEM & GEM into a final Color Edge Map (CEM).

HGC Edge Detector —

Performance comparison

• Compared methods– A grayscale edge detector (Sobel)– HGC edge detector– A YUV color edge detector

• Compared aspects– Speed– Edge completeness

• Testing data: real-life video images

HGC Edge Detector —

Speed comparison

• HGC edge detector saves average 20% of processing time compared to the YUV color edge detector.

HGC Edge Detector —

Comparison of edge completenessHGC Edge Detector —

Comparison of edge completeness (continued)

HGC Edge Detector —

Part 2: Edge Color Distribution Space

• Why introducing a Edge Color Distribution Space (ECDS) ?– 2D edge space is crowded.– Color is an important information to segment

different objects.

• Object discussed here is uniform-color object or textured object, not high-level object.

• The discussed image is of width W, of height H, and of 256-level grayscale.

Directional color operator• Get the directional average color of a point

• Edge point (x, y, g): 0xW, 0yH, 0g255

ECDS —

X-Y-G space ECDS

• Quantization– ECDS

– (x,y,g)(mx,my,gl)

• Distance-weighted accumulation

ECDS —

Characteristics of ECDS

• Spatial relation of an object in the image is kept.

• Objects of different colors are separated.

• The edge of uniform-color object is continuous.

• The edge of textured object is clustering.

ECDS —

ECDS: a synthetic imageECDS —

ECDS: a video imageECDS —

End.

Thank you!