Sobel edge detector

Question

Your task is to write a program that takes an input image and run it through edge-detection to become an output image.

The edge-detection works as follows (if unclear, see sobel edge detection):

• The value for a pixel is the total brightness of a pixel, so if it is in color, you will need to convert it to grayscale first (to keep things simple and golf-able, you can take the average value for R, G and B).
• The formulae for G_x and G_y for pixel p_(i,j) are:
  • G_x = -1 * p_{(i-1, j-1)} - 2 * p_{(i-1, j)} - 1 * p_{(i-1, j+1)} + 1 * p_{(i+1, j-1)} + 2 * p_{(i+1, j)} + 1 * p_{(i+1, j+1)}
  • G_y = -1 * p_{(i-1, j-1)} - 2 * p_{(i, j-1)} - 1 * p_{(i+1, j-1)} + 1 * p_{(i-1, j+1)} + 2 * p_{(i, j+1)} + 1 * p_{(i+1, j+1)}
• The value for the size of the edge at that pixel is then: √(G_x² + G_y²)

The output image is for each pixel the size of the edge √(G_x² + G_y²) as greyscale.

Bonuses:

• Perform a gaussian blur to smooth out the image before edge-detection kicks in, to omit any smaller edges. This gives a bonus of -30% on the end result.
• Take the angle of the edge in account. You give the output pixel some color, by taking the same greyscale value and adding color from a color wheel using the angle obtained from the formula arctan(G_y/G_x). This gives another bonus of -30% on the end result.

Rules:

• You may omit the value for the edgepixels, and set them to black, or you may use 0 for any pixel outside the image.
• Your ouput image must be in an image format that can be opened on most computers.
• Output must be written to disk or be pipeable to a file.
• Input is given as a commandline argument, in the form of a relative path to the image, or piped in from the commandline.
• This is code golf, so shortest code in bytes wins!

Answer 1

J, 166 164 161 154 150 144 143 bytes.

Not golfed too much; I mostly collapsed my longer implementation (see below), so there's probably lots of room for improvement. Uses BMP library. Saves result in file o. I handled edgepixels by only using full 3x3 cells, so the final image has width and height smaller by 2 pixels.

load'bmp'
S=:s,.0,.-s=:1 2 1
p=:([:*:[:+/[:,*)"2
'o'writebmp~256#.3#"0<.255<.%:(S&p+(|:S)&p)3 3,.;._3(3%~])+/"1(3#256)#:readbmp}:stdin''
exit''

Usage:

echo 'image.bmp' | jconsole golf.ijs

Expanded:

load 'bmp'

sobel1 =: 3 3 $ 1 0 _1 2 0 _2 1 0 _1
NB. transposed
sobel2 =: |: sobel1
NB. read image
image =: readbmp }: stdin''
NB. convert default representation to R,G,B arrays
rgbimage =: (3 # 256) #: image
NB. convert to grayscale
greyimage =: 3 %~ (+/"1) rgbimage
NB. 3x3 cells around each pixel
cells =: 3 3 ,.;._3 greyimage
NB. multiply 3x3 cell by 3x3 sobel, then sum all values in it
partial =: 4 : '+/"1 +/"1 x *"2 y'
NB. square partial (vertical and horizontal) results, sum and root
combine =: [: %: *:@[ + *:@]
NB. limit RGB values to 255
limit =: 255 <. ]
newimage =: limit (sobel1&partial combine sobel2&partial) cells
NB. convert back to J-friendly representation
to_save =: 256 #. 3 #"0 <. newimage
to_save writebmp 'out.bmp'
NB. jconsole stays open by default
exit''

Sample input and output:

Answer 2

Python, 161*0.7=112.7 bytes

With the Gaussian Blur bonus.

As you did not explicitly forbid built-in methods, here is OpenCV:

from cv2 import*
from numpy import*
g=GaussianBlur(cvtColor(imread(raw_input()),6),(3,3),sigmaX=1)
x,y=Sobel(g,5,1,0),Sobel(g,5,0,1)
imwrite('s.png',sqrt(x*x+y*y))

Without bonus, 136 bytes

from cv2 import*
from numpy import*
g=cvtColor(imread(raw_input()),6)
x,y=Sobel(g,5,1,0),Sobel(g,5,0,1)
imwrite('s.png',sqrt(x*x+y*y))

• Edit1: Replaced the named constans by their values.
• Edit2: Uploaded samples

Answer 3

MATLAB, 212*0.4=84.8 bytes

Using the filter toolbox and the HSV colorspace

function f(x);f=@(i,x)imfilter(i,x);s=@(x)fspecial(x);S=s('sobel');A=f(double(rgb2gray(imread(x)))/255,s('gaussian'));X=f(A,S);Y=f(A,S');imwrite(hsv2rgb(cat(3,atan2(Y,X)/pi/2+0.5,0*A+1,sqrt(X.^2+Y.^2))),'t.png')

or ungolfed

function f(x)
f=@(i,x)imfilter(i,x);
s=@(x)fspecial(x);
S=s('sobel');
A=f(double(rgb2gray(imread(x)))/255,s('gaussian'));
X=f(A,S);
Y=f(A,S');
imwrite(hsv2rgb(cat(3,atan2(Y,X)/pi/2+0.5,0*A+1,sqrt(X.^2+Y.^2))),'t.png')

Answer 4

Love2D Lua, 466 Bytes

A=arg[2]i=love.image.newImageData q=math t=i(A)g=i(t:getWidth()-2,t:getHeight()-2)m={{-1,-2,-1},{0,0,0},{1,2,1}}M={{-1,0,1},{-2,0,2},{-1,0,1}}t:mapPixel(function(_,_,r,g,b)a=(r+g+b)/3 return a,a,a end)g:mapPixel(function(x,y)v=0 for Y=0,2 do for X=0,2 do v=v+(t:getPixel(x+X,y+Y)*m[Y+1][X+1])end end V=0 for Y=0,2 do for X=0,2 do V=V+(t:getPixel(x+X,y+Y)*M[Y+1][X+1])end end v=q.max(q.min(q.sqrt(V^2+v^2),255),0)return v,v,v end)g:encode('png',"o")love.event.quit()

Takes command line input, outputs to a file called "o" under your Love2D appsdata folder. Love2D Wont let you save files anywhere else.

Just about as golfed as I could get it, probably could be golfed further.

Explained

-- Assign the Input to A
A=arg[2]


-- Assign some macros to save FUTURE BYTES™
i=love.image.newImageData
q=math

-- t is the original image, g is the new output image. g is two pixels smaller, which is easier and better looking than a border.
t = i(A)
g = i(t:getWidth()-2,t:getHeight()-2)

-- m and M are our two sobel kernals. Fairly self explanitary.
m = {{-1,-2,-1}
    ,{0,0,0}
    ,{1,2,1}}

M = {{-1,0,1}
    ,{-2,0,2}
    ,{-1,0,1}}

-- Convert t to grayscale, to save doing this math later.
t:mapPixel(function(_,_,r,g,b)a=(r+g+b)/3 return a,a,a end)

-- Execute our kernals
g:mapPixel(function(x,y)
    -- v refers to the VERTICAL output of the Kernel m.
    v=0
    for Y=0,2 do
        for X=0,2 do
            v=v+(t:getPixel(x+X,y+Y)*m[Y+1][X+1])
        end
    end

    -- V is the HORIZONTAL of M
    V=0
    for Y=0,2 do
        for X=0,2 do
            V=V+(t:getPixel(x+X,y+Y)*M[Y+1][X+1])
        end
    end

    -- Clamp the values and sum them.
    v = q.max(q.min(q.sqrt(V^2 + v^2),255),0)
    -- Return the grayscale.
    return v,v,v
end)

-- Save, renaming the file. The golfed version just outputs as 'o'
g:encode('png',"S_".. A:gsub("(.*)%....","%1.png"))

-- Quit. Not needed, but I'm a sucker for self contained LOVE2D
love.event.quit()

Test

And...

Although it doesn't actually improve my score (Makes it worse infact), here is the version with the colour wheel implemented.

900 - 270 = 630 Bytes

A=arg[2]i=love.image.newImageData q=math t=i(A)g=i(t:getWidth()-2,t:getHeight()-2)m={{-1,-2,-1},{0,0,0},{1,2,1}}M={{-1,0,1},{-2,0,2},{-1,0,1}}function T(h,s,v)if s <=0 then return v,v,v end h,s,v=h*6,s,v/255 local c=v*s local x=(1-q.abs((h%2)-1))*c local m,r,g,b=(v-c),0,0,0 if h < 1 then r,g,b=c,x,0 elseif h < 2 then r,g,b=x,c,0 elseif h < 3 then r,g,b=0,c,x elseif h < 4 then r,g,b=0,x,c elseif h < 5 then r,g,b=x,0,c else r,g,b=c,0,x end return(r+m)*255,(g+m)*255,(b+m)*255 end t:mapPixel(function(_,_,r,g,b)a=(r+g+b)/3 return a,a,a end)g:mapPixel(function(x,y)v=0 for Y=0,2 do for X=0,2 do v=v+(t:getPixel(x+X,y+Y)*m[Y+1][X+1])end end V=0 for Y=0,2 do for X=0,2 do V=V+(t:getPixel(x+X,y+Y)*M[Y+1][X+1])end end h=v H=V v=q.max(q.min(q.sqrt(V^2+v^2),255),0)h=q.atan2(H,h)/q.pi*2 return T(h,1,v,255)end)g:encode('png',"S_".. A:gsub("(.*)%....","%1.png"))G=love.graphics.newImage(g)love.event.quit()