Image Processingon the GPUEdge Detection

(with implementation on a GPU)And

Text Recognition(if time permits)

Jared BarnesChris Jackson

Edge Detection◦ Wikipedia: Identifying points in a digital image at

which the image has discontinuities.

Edge Detection

http://4.bp.blogspot.com/-p_9w91wC_Rc/TbBgF7dQYhI/AAAAAAAAACM/DQTrM_a7Apg/s1600/edge-example-c.png

Canny Edge Detector John Canny “A Computational Approach to Edge

Detection” 1986

http://ieeexplore.ieee.org.libproxy.mst.edu/stamp/stamp.jsp?tp=&arnumber=4767851

1. Noise Removal

2. Image Gradient Computation

3. Non-Maximum Suppression

4. Hysteresis Thresholding

Canny Edge Detector

Gaussian Smoothing or Blurring

A pixel is changed based on a weighted average of itself and its neighbors

The number of neighbors (3x3, 5x5) and the relative weights can vary

Noise Removal

3D Gaussian Distribution

Normalized 2D Gaussian Approximation

http://www.librow.com/content/common/images/articles/article-9/2d_distribution.gif

http://homepage.cs.uiowa.edu/~cwyman/classes/spring08-22C251/homework/canny.pdf

Gaussian Smoothing – Avoid Blur

Too much

About right

http://media.tumblr.com/ccd6945141b46e5e2f5c36168f6a8037/tumblr_inline_mhcv1l0EZB1qz4rgp.png

http://www.eversparkinteractive.com/wp-content/uploads/2013/03/gaussian-blur-thumbnail.jpg

Spotty Smooth

Calculus!

First derivative inthe X and Y directions(separately)

Gradient Computation

http://homepage.cs.uiowa.edu/~cwyman/classes/spring08-22C251/homework/canny.pdf

Gx Gy

Sobel Operator (2 kernels)

Then round to:0° =←→ 90°=↑↓45°=↗↙ 135°=↘↖

Image Gradient: Sobel Operator

http://suraj.lums.edu.pk/~cs436a02/CannyImplementation.htm

X Gradient

(Horizontal Edges)

Y Gradient(Vertical Edges)

Make edges exactly one pixel thick Look at the gradient magnitude of your 2

neighbors in the direction of your angle

Non-Maximum Suppression

80 85 90

80 85 90

80 85 90

35 35 50

35 50 40

50 40 40

Example 2Angle = 0° ←→

Example 1Angle = 135° ↘↖

80 85 90

80 85 90

80 85 90

35 35 50

35 50 40

50 40 40

80 85 90

80 0 90

80 85 90

35 35 50

35 50 40

50 40 40

Keep it!

Kill it!

Non-Maximum Suppression

http://suraj.lums.edu.pk/~cs436a02/CannyImplementation.htm

Thick Edges

(Gradient Magnitude)

Thin Edges

(Gradient Magnitude)

Two thresholds are better than one!

If a pixel’s value is above Thigh, it’s an edge.

If a pixel’s value is below Tlow, it’s not an edge.

If a pixel’s value is between Thigh and Tlow, it might be an edge (provided it is connected to an actual edge)

Hysteresis Thresholding

0 0 0 0 0 0 0

0 50 50 50 50 50 0

0 0 0 50 0 0 0

0 0 0 50 0 0 0

0 36 0 50 0 40 0

39 0 0 50 0 0 39

0 38 0 50 0 43 0

0 0 0 50 0 0 0

0 40 40 50 40 40 0

0 0 0 0 0 0 0

Thigh = 45 Tlow = 35

0 0 0 0 0 0 0

0 50 50 50 50 50 0

0 0 0 50 0 0 0

0 0 0 50 0 0 0

0 36 0 50 0 40 0

39 0 0 50 0 0 39

0 38 0 50 0 43 0

0 0 0 50 0 0 0

0 40 40 50 40 40 0

0 0 0 0 0 0 0

0 0 0 0 0 0 0

0 50 50 50 50 50 0

0 0 0 50 0 0 0

0 0 0 50 0 0 0

0 36 0 50 0 40 0

39 0 0 50 0 0 39

0 38 0 50 0 43 0

0 0 0 50 0 0 0

0 40 40 50 40 40 0

0 0 0 0 0 0 0

0 0 0 0 0 0 0

0 255 255 255 255 255 0

0 0 0 255 0 0 0

0 0 0 255 0 0 0

0 0 0 255 0 0 0

0 0 0 255 0 0 0

0 0 0 255 0 0 0

0 0 0 255 0 0 0

0 255 255 255 255 255 0

0 0 0 0 0 0 0

1. Smooth image to reduce noise

2. Calculate X & Y derivatives to get edges

3. Thin all edge widths to 1 pixel

4. Remove weak, unconnected edges

(ta da!)

Canny Edge Detector

How do we parallelize the Canny Edge Detector?

Parallelization!

Convolution – Independent of order

Parallelize: Canny Edge Detector

5 5510 101020 2020

2 224 222 22

Image Kernel2 224 222 22

10 101040 202040 4040

Element-wise Multiplication

230Sum All Values

11Divide by Kernel

Sum

11

Convolve a Gaussian Kernel with the image

Each GPU core can convolve each pixel in the image individually with the Gaussian Kernel

One thread per pixel, each performing 9 multiplies, 9 adds, and 1 division

Embarrassingly Parallel with huge speedup

Step 1: Gaussian Smoothing

Convolve two Sobel Kernels with the image

Wait, convolution again?

Same as previous step – we can even reuse the convolution function!

Step 2: Gradient Computation

Comparing 3 pixel gradient magnitudes and clearing the middle pixel or leaving it alone

Similar to convolution… but simpler! Each GPU thread owns a pixel:

1. Check gradient angle of pixel2. Compare this pixel’s magnitude with two

neighbors in the direction of its angle3. If I’m greater than those neighbors, leave me

alone; otherwise, mark me as “not an edge” Less speedup than steps 1 and 2

Step 3: Non-Maximum Suppression

Mark pixels > Thigh as strong edges Mark pixels < Tlow as not edges Mark remaining pixels as weak edges if they

connect to a strong edge Typically implemented with recursion Each thread with a weak-edge pixel looks at

nearest 2 neighbors to find a strong-edge pixel

With identical algorithms on CPU and GPU, speedup is marginal (memory accesses, not much processing)

Step 4: Hysteresis Thresholding

Optical Character Recognition

http://www.flacom.com/content/uploads/2013/09/hello-world.jpg

Wikipedia: The mechanical or electronic conversion of images of printed text into computer-readable text.

http://hackadaycom.files.wordpress.com/2010/09/helloworldconsole.png

Label Connected Components

Look For Letters

Adjust for disconnected letters

I have the edges – Now what?

HELLO WORL

D

HELLO WORL

D

EF

? ü j i

Create a list of components in the image A component is simply a set of connected

edges1. Label each edge pixel with a unique

component ID2. Examine each pixel’s 8 touching neighbors

and set that pixel’s ID to the smallest neighbor ID

3. Repeat step 2 until no pixel IDs are changed

Connected Components Labelling

Uhh… what’s a letter? How do we know it’s a letter? How does the computer know it’s a letter?

Look For Letters

Letters are represented by a vector of numbers indicating the ratio of black pixels to white pixels in each division of the letter-image.

Wait, what is a letter?

A0 40 0 00

60 40 60 5555 15 55 0010 75 10 00

15 0 15 1515

Compute how closely each labelled component matches each letter in your alphabet

The component is then marked with whichever letter it most closely matches

Okay – Now look for letters

Letters like ‘i’ and ‘j’ have floating parts Sometimes edge detection may accidentally

break up a letter A letter vector should then get an additional

property indicating vertical discontinuity

Correct for Disjoint Letters

T T E R V EL E O R … … …C T 0/1