Pages

.

Sobel edge detection using Java Advanced Imaging



The Java Advanced Imaging API supports a number of interesting convolutions straight out of the box, and one of them is Sobel edge detection.

The Sobel edge-detection kernel comes in two varieties, corresponding to horizontal edge detection and vertical edge detection:
 1  2  1
0 0 0
-1 -2 -1

1 0 -1
2 0 -2
1 0 -1
You can combine them, and/or run them serially against an image, to detect all edges in an image. And that's what the following code example (in JavaScript) does. You can run the following script against an image of your choice using the ImageMunger app I wrote about a few days ago. Be sure the Java Advanced Imaging JARs are in your classpath.
/* Sobel.js
* Kas Thomas
* 31 January 2010
* Public domain.
*
* An edge-detection routine using
* Java Advanced Imaging.
*
* Requires Java Advanced Imaging library:
* http://java.sun.com/products/java-media/jai/current.html
*
* Run this file using ImageMunger:
* http://asserttrue.blogspot.com/2010/01/simple-java-class-for-running-scripts.html
*
*/


jai = Packages.javax.media.jai;
sobelH = jai.KernelJAI.GRADIENT_MASK_SOBEL_HORIZONTAL;
sobelV = jai.KernelJAI.GRADIENT_MASK_SOBEL_VERTICAL;

pb = new Packages.java.awt.image.renderable.ParameterBlock( );

// ImageMunger puts "Image" in global scope:
pb.addSource( Image );
pb.add( sobelH );
pb.add( sobelV );

renderedOp = jai.JAI.create( "gradientmagnitude", pb );
var image = renderedOp.getRendering().getAsBufferedImage();
Panel.setImage( invertImage( image ) );

// take BufferedImage as arg; flip all bits in all pixels
function invertImage( image ) {

var w = image.getWidth();
var h = image.getHeight();
var pixels = image.getRGB( 0,0, w,h, null, 0,w );

for ( var i = 0; i < pixels.length; i++ )
pixels[ i ] =~ pixels[ i ]; // flip pixel bits

image.setRGB( 0,0, w,h, pixels, 0, w );
return image;
}

If you run the Sobel operation by itself, you get a "negative" image, like so:



If you want the inverted version of this image (see example further above), you need to invert the individual pixels of the image. The super-fast way to do it is with a lookup table, but you can also do the pixel inversion manually, in a loop, which is what I've done (for illustrative purposes). In JavaScript the pixel-inversion loop adds about one second of processing time for a 600 x 400 image. The Sobel filter by itself takes around a half a second.

Sobel tends to be very sensitive to noise (it will feature-enhance specks and JPEG artifacts), so it often helps to smooth an image, first, with a blurring operation, prior to applying Sobel.

Future projects:
  • Write a "tunable" version of Sobel that can detect soft or hard edges, according to a tuning parameter.
  • Write a version of Sobel that's tunable by color (viz., detecting just blue edges, or just black edges, or just medium-grey edges).
reade more... Résuméabuiyad

How to implement custom Paint in 50 lines of Java



Java offers many ways to customize strokes and fills, including the use of gradient fills and image fills (see this excellent tutorial by Marty Hall), but we tend to forget that "procedural textures" are easily implemented as custom Paint.

The text shown above was painted using a custom Paint class, SinePaint.java, consisting of around 50 lines of code (not counting imports), as shown below. (Scroll the code sideways to see lines that didn't wrap.) The SinePaint class procedurally generates a sine-wave fill pattern in the red color channel, running in the 'y' direction (vertical sine wave).

/* SinePaint
* Kas Thomas
* 30 January 2010
* Public domain.
* http://asserttrue.blogspot.com/
*
* A quick example of how to implement java.awt.Paint
*/


import java.awt.Paint;
import java.awt.PaintContext;
import java.awt.Rectangle;
import java.awt.RenderingHints;
import java.awt.geom.AffineTransform;
import java.awt.geom.Rectangle2D;
import java.awt.image.ColorModel;
import java.awt.image.Raster;
import java.awt.image.WritableRaster;

class SinePaint implements Paint {

public SinePaint() {}

public PaintContext createContext(ColorModel cm,
Rectangle deviceBounds,
Rectangle2D userBounds,
AffineTransform xform,
RenderingHints hints) {
return new Context(cm, xform);
}

public int getTransparency() {
return java.awt.Transparency.OPAQUE;
}

class Context implements PaintContext {

public Context(ColorModel cm_, AffineTransform xform_) { }

public void dispose() {}

public ColorModel getColorModel() {
return ColorModel.getRGBdefault();
}

public Raster getRaster(int xOffset, int yOffset, int w, int h) {

WritableRaster raster =
getColorModel().createCompatibleWritableRaster(w, h);
float [] color = new float[4];

// Row major traversal.
for (int j = 0; j < h; j++) {
for (int i = 0; i < w; i++) {
color[3] = 255;
color[2] = color[1] = 0;
// Write a sine-wave pattern to the Red channel
color[ 0 ] =
(1 + (float) Math.sin( 6.28f*((double) j)/h )) * .5f * 255;;

raster.setPixel(i, j, color);
} // i
} // j
return raster;
} // getRaster()
} // Context
} // SinePaint

Implementing the Paint interface turns out not to be such a big deal. There's only one required method, createContext():
   public PaintContext createContext(ColorModel cm,
Rectangle deviceBounds,
Rectangle2D userBounds,
AffineTransform xform,
RenderingHints hints)
Most of the formal parameters are hints and can safely be ignored. Note that this method returns a java.awt.PaintContext object. It turns out PaintContext is an interface as well, so you do end up having to implement it, and this is where the real action occurs. The methods of the PaintContext interface include:

public void dispose() {};
public ColorModel getColorModel();
public Raster getRaster(int x,
int y,
int w,
int h);

The dispose() method releases any resources that were allocated by the class. In our case, we allocated nothing and so our dispose method is empty. The getColorModel() method can, in most cases, be a one-liner that simply returns ColorModel.getRGBdefault(). The real action is in getRaster(). That's where you have the opportunity to set the pixel values for all the pixels in the raster based on their x-y values. If you're familiar with shaders and/or procedural textures, you know what this is about. This is your opportunity to shade an area in accordance with a pixel's x-y location onscreen (or rather, within the image).

If you've been using the ImageMunger app I wrote about a few days ago, you can run the following script with it to see SinePaint in operation. (This is the script that produced the colored text shown above.)

/* paintedText.js
* Kas Thomas
* 30 January 2010
* Public domain.
*
* Run this file using ImageMunger:
* http://asserttrue.blogspot.com/2010/01/simple-java-class-for-running-scripts.html
*/


g2d = Image.createGraphics();

rh = java.awt.RenderingHints;
hint = new rh( rh.KEY_TEXT_ANTIALIASING,rh.VALUE_TEXT_ANTIALIAS_ON );
g2d.setRenderingHints( hint );

sinePaint = new Packages.SinePaint( );

g2d.setPaint( sinePaint );
g2d.setFont( new java.awt.Font("Times New Roman",java.awt.Font.BOLD,130) );
g2d.drawString( "Shiny",50,100);
g2d.drawString( "Text",50,200);

Panel.updatePanel();
(Scroll sideways to see lines that didn't wrap.)

Future projects:
  • Make use of the AffineTransform argument to enable scaled and otherwise transformed textures.
  • Implement Perlin noise as a type of Paint.
  • Implement "bump map" 3D effects in Paint.
reade more... Résuméabuiyad

Image rotation in 8 lines using the Java Advanced Imaging API


Last time, I showed that you could rotate an image 180 degrees in what amounts to one line of JavaScript, basically just doing pixels.reverse( ). Rotating an image by an arbitrary amount (something other than 180 degrees) is almost as easy. It requires about 8 lines of code.

JavaScript for doing the rotation with the Java Advanced Imaging API is shown below. JAI makes short work of this and a ton of other graphics transformations. All you have to do is be sure the JAI JARs are in your classpath. Then you can set up a transformation by creating a ParameterBlock with appropriate parameters (in this case, the x- and y-coordinates of the rotation origin, the amount of rotation in radians, and optionally a rendering hint as to what kind of pixel interpolation you'd like; in this case, we don't specify a hint and thus accept the default of INTERP_NEAREST).

/*
* rotate.js
* Kas Thomas
* 29 January 2010
* Public domain.
*
* Requires Java Advanced Imaging library:
* http://java.sun.com/products/java-media/jai/current.html
*
* Run this file using ImageMunger:
* http://asserttrue.blogspot.com/2010/01/simple-java-class-for-running-scripts.html
*/


pb = new Packages.java.awt.image.renderable.ParameterBlock( );
pb.addSource( Image );

pb.add( new java.lang.Float(0) ); // x-origin
pb.add( new java.lang.Float(0) ); // y-origin
pb.add( new java.lang.Float( Math.PI/8) ); // rotation amount

renderedOp = Packages.javax.media.jai.JAI.create( "rotate", pb );
image = renderedOp.getRendering().getAsBufferedImage();

Panel.setImage( image );

Note that JAI expects parameters of type float, which is not what JavaScript provides by default. By default, numbers in JavaScript are doubles. So you have to explicitly create java.lang.Floats as shown.

As in previous posts, I'm using my little ImageMunger app to run the script.

Also, as with previous scripts, performance is quite good (through no fault of my own): rotation occurs at a rate of about 500 pixels per millisecond on a Dell Inspiron laptop with 2.2 GHz Intel Duo processor running (gack!) Windows Vista. Which ain't bad at all. I'll take 500-pixels-per-millisec throughput any day, on any OS.
reade more... Résuméabuiyad

Fast image rotation using JavaScript


Five lines of JavaScript suffice to rotate the image on the left 180 degrees, at a processing speed of one million pixels per second.

Rotating an image 180 degrees turns out to be extremely easy -- and as good an illustration as any that JavaScript doesn't automatically have to mean "slow." The test image shown here (a 600 x 446-pixel JPEG) was rotated 180 degrees using the script below in only 262 milliseconds -- quite literally the blink of an eye. (Note that to use this script, you need the small Java app -- ImageMunger -- that I wrote about a few days ago.)

w = Image.getWidth();
h = Image.getHeight();
pixels = Image.getRGB( 0,0, w,h, null, 0, w );
Image.setRGB( 0,0, w,h, pixels.reverse(), 0, w );
Panel.updatePanel( );
The key to what's going on is the call to pixels.reverse(). When you call getRGB() on a BufferedImage, Java hands you back all the image's pixels in a one-dimensional array. Simply rewriting the array in reverse order has the effect of causing the image to paint lower-right-corner-first (if you know what I mean). It rotates the image 180 degrees.

The reverse() method of the Array object (built into JavaScript) is implemented in bytecode and runs at bytecode speed (as is true of all native methods). It means that the "main loop" (the pixel-reversal code) runs at compiled-code speed.

The real lesson here is that if you want your JavaScript code to run fast, you should take advantage, whenever you can, of native methods of the language. Those methods are nearly always implemented as hand-optimized C or Java -- as I discussed in an earlier post. It's free speed. Don't let it go to waste.
reade more... Résuméabuiyad

HTTPbis

In case you haven't been following developments around things like HTTPbis and the PATCH verb, take a look at this slideshow. HTTP is far from a static specification. It's about to undergo significant clarification and expansion. All I can say is: It's about time.
reade more... Résuméabuiyad

Sharpening an image using java.awt.image.ConvolveOp



The somewhat blurry image on the left was sharpened, to produce the image on the right, using the two dozen or so lines of JavaScript code shown further below (run in conjunction with the ImageMunger Java app; see text for details).


As with contrast adjustment, sharpening an image can be thought of as an exercise in weak signal amplification. Generally it means making the differences between neighboring pixels more noticeable. You can do this by brute-force analysis of pixels, of course, but area operators -- kernel-based convolutions -- are the clean way to go.

A convolution kernel is a 2D matrix of numbers that can be used as coefficients for numerical operations on pixels. Suppose you have a 3x3 kernel that looks like this:
1 2 1
2 0 2
1 2 1
As you loop over all pixels in the image, you would, for any given pixel, multiply the pixel itself by zero; multiply the pixel directly above the given pixel by 2; also multiply by 2 the pixels to the left, right, and below the pixel in question; multiply by one the pixels at 2 o'clock, 4 o'clock, 8 o'clock, and 10 o'clock to the pixel in question; add all these numeric values together; and divide by 9 (the kernel size). The result is the new pixel value for the given pixel. Repeat for each pixel in the image.

In the example just cited, the kernel (1 2 1 etc.) would end up smoothing or blurring the image, because in essence we are replacing a given pixel's value with a weighted average of surrounding pixel values. To sharpen an image, you'd want to use a kernel that takes the differences of pixels. For example:
 0 -1  0
-1 5 -1
0 -1 0
This kernel would achieve a differencing between the center pixel and pixels immediately to the north, south, east, and west. It would cause a fairly harsh, small-radius (high frequency) sharpening-up of image features.

It turns out, Java has good support for kernel-based 2D convolutions of images using java.awt.image.Kernel and java.awt.image.ConvolveOp. It takes only a few lines of JavaScript to run a convolution kernel against a JPEG (or other image) to achieve sharpening of the image; see code listing below. (A few days ago, I posted code for a small Java app -- called ImageMunger -- that opens an image of your choice and runs a script against it. You may want to use that app to run the following code.)

kernel = [ .25, -2, .25,
-2, 10, -2,
.25, -2, .25 ];

function normalizeKernel( ar ) {

for (var i = 0, n = 0; i < ar.length; i++)
n += ar[i];
for (var i = 0; i < ar.length; i++)
ar[i] /= n;

return ar;
}

kernel = normalizeKernel( kernel );

k = new java.awt.image.Kernel( 3,3,kernel );
convolver =
new
java.awt.image.ConvolveOp( k,
java.awt.image.ConvolveOp.EDGE_NO_OP,
null);


target =
new java.awt.image.BufferedImage( Image.getWidth(),
Image.getHeight(),Image.getType() );

g = target.createGraphics( );
g.drawImage( Image, null,0,0 );
g.dispose();
Image = convolver.filter( target, Image );
Panel.updatePanel( );
Recall that the ImageMunger app I talked about here a few days ago exports a couple of global variables to the JavaScript context: namely, Image (a handle to the BufferedImage) and Panel (a reference to the JComponent in which the image is being displayed). With the aid of those globals and appropriate calls to JRE methods, it's very easy to run a convolution. Easy and fast: Expect to process around a thousand pixels per millisecond.

Future projects:
  • Programmatically generate and initialize large kernels. Have a slider-based UI that performs interesting initializations of kernel values.
  • Kernel-based sharpening tends to preferentially add high frequencies to an image, which can be problematic in images that have lots of areas of high-frequency noise. Create a "smart sharpen" algorithm that dynamically tunes the frequency of the sharpening (kernel values) according to the natural "humm" (the natural frequencies) of the area or areas that are being sharpened.
  • As a side benefit of the foregoing, create a sharpening algorithm that won't sharpen JPEG artifacts.
reade more... Résuméabuiyad

Fast contrast adjustment using Perlin's gain function


Lena (a standard test image in the graphics programming world) is transformed by the Perlin gain function with b = 0.5, b = 0.4, and b = 0.65, respectively (left to right).

Coaxing more contrast out of an image is a kind of exercise in weak-signal detection. You're trying to make information more obvious in the presence of noise. In practice, it comes down to making light tones lighter and dark ones darker. But you have to do it carefully, in such a way as not to force too many (preferably no) dark tones into total blackness nor too many light tones into total whiteness.

A very useful approach is to adapt Ken Perlin's gain function to the remapping of pixel values. In JavaScript, the function looks like this:

var LOG_POINTFIVE = -0.6931471805599453;

function gain( a, b) {

var p = Math.log(1. - b) / LOG_POINTFIVE;

if (a < .001)
return 0.;
if (a > .999)
return 1.;
if (a < 0.5)
return Math.pow(2 * a, p) / 2;
return 1. - Math.pow(2 * (1. - a), p) / 2;
}
The Perlin gain function maps the unit interval (i.e., real numbers from 0 to 1, inclusive) onto itself in such a way that a control value of 0.5 maps [0..1] to [0..1] unchanged, whereas a control-parameter value of (say) 0.25 maps the number 0.5 to itself but maps all other values in the unit interval to new values, as shown in the figures below. (In each graph, the x- and y-axes go from zero to 1.0.)

In the function shown above, the control parameter (the formal parameter that throttles the function) is b. The input value that will be mapped to a different output is a. If b is 0.5, then the output of the function will always be a no matter what value of a you pass in. But if b is 0.25, and a is (say) .4, the function will return 0.456, whereas if b is 0.75 and a is 0.4, the output will be 0.32. In one case, a got bigger, and in the other case it got smaller. The function has the effect of making bigger values bigger and smaller values smaller when b > 0.5. It has the effect of making bigger values smaller and smaller values bigger when b < 0.5.


Gain = 0.25


Gain = 0.5


Gain = 0.75

This turns out to be a great function for changing the contrast of an image. All you have to do is re-map pixel values onto the [0..255] interval using the function, with a value for b of something greater than 0.5 if you want the image to have more contrast, or less than 0.5 if you want the image to have less contrast.

It turns out Java.awt.image has a built-in class called LookupOp that implements a lookup operation from a source color table to an output color, which makes it easy to implement extremely high performance contrast adjustment via the gain function. The red, green, and blue values in an RGB image span the interval zero to 255. All we need to do is create a byte array of length 256, containing those values, then modify the table by passing each value through the gain function. The altered table can be used to create a LookupOp instance, and then you just need to call filter() on the instance, passing it an input image and an output (holder) image.

I do all this in JavaScript in the code listing below. To run this script against an image of your choice, you simply need the (open source) ImageMunger app that I wrote about a couple days ago.

/* Contrast.js
Kas Thomas
24 Jan 2010

Public domain.

http://asserttrue.blogspot.com/
*/



// Use values >.5 but <>
// < .5 && > 0 for less contrast.
CONTRAST = .75; // Adjust to taste!

awtImage = java.awt.image;

/* Return a java.awt.image.ByteLookupTable */
function getLUT( amt ) {

var LOG_POINTFIVE = -0.6931471805599453;

function gain( a, b) {

var p = Math.log(1. - b) / LOG_POINTFIVE;

if (a < .001)
return 0.;
if (a > .999)
return 1.;
if (a < 0.5)
return Math.pow(2 * a, p) / 2;

return 1. - Math.pow(2 * (1. - a), p) / 2;
}

// Perlin's gain function
// per K. Perlin, "An Image Synthesizer,"
// Computer Graphics, v19, n3, p183-190 (1985)

/* We are going to construct a table of values,
0..255, wherein the values vary nonlinearly
according to the formula in gain(), as
throttled by the parameter 'amt' */


var tableSize = 256;
var javaArray = java.lang.reflect.Array;
var bytes =
javaArray.newInstance( java.lang.Byte.TYPE,
tableSize );
var lut = new awtImage.ByteLookupTable(0, bytes);

for (var i = 0,gainValue = 0; i < tableSize; i++) {
gainValue = gain(i / 255., amt);
var byteValue = 255 & (255. * gainValue);
if (byteValue >= 128)
byteValue = -(255 - byteValue);
bytes[i] = byteValue;
}

return lut;
}


// Create the lookup table
lut = getLUT( CONTRAST );

// Create the java.awt.image.LookupOp
lop = new awtImage.LookupOp( lut, null );

// Clone the source image
src = theImage;
clone = new awtImage.BufferedImage( src.getWidth(),
src.getHeight(), src.getType() );
g2d = clone.getGraphics();
g2d.drawImage( src, 0,0,null );

// apply the contrast
lop.filter( clone, src );

// refresh the screen
//Panel.updatePanel();
thePanel.repaint();
There are a couple of things to note. First, you can't assume that an array created in JavaScript can be used as an array in a Java context, because Java is type-fussy and JavaScript isn't, so instead, to create a Java array in JavaScript you have to do:

var bytes =
javaArray.newInstance( java.lang.Byte.TYPE,
tableSize );
This creates a Java byte array and stores a reference to it in JavaScript. But now you have another problem, which is that you can't just map values from [0..255] onto a 256-length Java byte array, because in Java, the byte type cannot accommodate values greater than 127. In other words, byte is a signed 8-bit type, and you'll get an error if you try to store a value of 128 in a byte variable. So to initialize the 256-length byte array with values from zero to 0xFF, and still keep the Java compiler happy, we have to resort to a bit of twos-complement legerdemain:

 if (byteValue >= 128)
byteValue = -(255 - byteValue);
To test the contrast.js script, I ran it against the Lena image with values for b -- in the gain function -- of 0.5 (which leaves the image unchanged), 0.4 (which reduces contrast), and 0.65 (which increases contrast). You can see the results at the top of this post. The code executes very quickly because it's all a table lookup done in Java. The JavaScript part simply initializes a (very small) table.

Projects for the future:
  • Add a UI to the program to allow contrast to be adjusted in real time with a slider.
  • When contrast is increased in a color image, hot parts of the image appear to get hotter and cold parts appear to get colder. Write a modification of the contrast code that compensates for apparent temperature changes.
  • In the code as currently written, inaccuracies due to rounding errors are ignored. Rewrite the routine to transform the image to a 16-bit color space, and use a 32K color lookup table rather than a 256-byte table, for the contrast adjustment; then transform the image back to its original color space afterwards.
reade more... Résuméabuiyad

Fast conversion of bitmaps to SVG


The image on the left is the 600x446-pixel "original" RGB image; the copy on the right has been tiled into 6706 solid-filled rectangles. See text for discussion.

Conversion of bitmapped images (*.jpg, *.gif, etc.) to vector format (Postscript, SVG) is, in general, a Difficult Problem. In some ways, it's not unlike trying to convert spoken text (in the form of *.mp3 audio) to ASCII. Well okay, maybe it's not that bad, but it's gnarly. You're trying to parse precise geometric shapes out of an ensemble of intensity samples. It's not at all obvious how to do it. Finding a fully general algorithm for describing an arbitrary bitmap as an ensemble of vector-drawable shapes is vexingly difficult. You might as well try to reassemble Yugoslavia.

You have to admit, though, it'd be darned handy to be able to transcode a bitmap image into (say) some flavor of XML (such as SVG). Unlike binary formats, XML is wire-protocol-friendly, queryable at the element and attribute level, transformable using XSLT (and/or E4X, or other technologies), human-readable (bwahhh-ha-ha . . .), and highly compressible. Converting JPEG to XML would open the door (potentially) to many interesting operations. Imagine running XPath queries against images to find specific areas of interest, or morphing one image into another using XSLT.

While there is obviously no one right way to proceed, I'd like to propose an approach to the problem (the problem of how to transcode bitmaps to SVG) based on quadtree decomposition of an image into polygons -- rectangles, in particular. It's possible to apply the same approach using triangles as primitives, instead of rects, or spline patches for that matter, but rectangles offer some noteworthy advantages. Most images are rectangles to begin with, of course, and parsing them into smaller (but adjoining) rects is a natural thing to do. Once parsed, rects allow a relatively compact representation; and at render-time, not many solid shapes are easier or faster to draw.

We have the considerable advantage of being able to start from the perspective of seeing a bitmapped image as just a grid of one-by-one rectangles (known as pixels). Two adjoining pixels of the same color become a 1x2 solid-color rect, four adjoining identical pixels become a 2x2 or 1x4 rect, and so on. Surely we can parse out the "naturally occurring" solid-filled rects within a pixel lattice? Those rects are instantly vector-drawable.

If we do this, we'll get what we get -- probably a very small handful of accidental finds. But we can do better, if (through preprocessing) we quantize pixel values in such a way as to force nearly-equal pixels to be equal in value. This will cause the appearance of a higher percentage of solid-color rects within a bitplane. Of course, when we're done, there's still no guarantee we will have covered the entire visual plane with rects (unless, as I say, we count individual pixels as 1x1 rects). Nevertheless, it's an interesting strategy: Force adjoining almost-equal pixels to be equal in value, then aggregate them into solid-color rects. Deal with the leftovers later.

At the extreme, we can look at the entire image as being a collection of pixels that are nearly equal in value. There is actually considerable variance in pixels, of course (unless the image is truly not very interesting). But what we might do is quantify the variance in some way, and if it exceeds a threshold, divide the image into daughter rects -- and begin again. We know that eventually, if we keep subdividing, we'll get to individual pixels (1x1 rects) that have zero variance within themselves. By this sort of strained logic we might convince ourselves that there should be less variance in small rects than in large ones, as a general trend.

Let me get right to it and propose an algorithm. (Listen up.) Start by considering a rectangle covering the whole image. Going pixel by pixel, calculate some variance measure (such as root-mean-square variation from average) for all pixels, for the whole region. If the variance is low (lower than some arbitrary limit), consider all pixels within the region to be equal; define the region as a rect of color Argb (representing the arithmetic average of the pixel values). If the variance exceeds an arbitrary threshold, subdivide the image. Specifically, subdivide it into four (generally unequal) rects.

To determine where to subdivide, calculate the "center of gravity" of the parent rect. A pixel's lightness or darkness, multiplied by its position vector, gives its moment-in-X and moment-in-Y. Add all the X moments together (and the Y moments). Divide by the number of pixels. The result is the visual center of the region. That's where you divide.

Repeat the procedure on newly created rectangular regions. Any regions with low variance can be rendered as a solid-color rect; the other regions are subdivided, and the algorithm continues until reaching individual pixels, or until every region larger than a pixel was successfully encoded as a rect.

This general technique is known as quadtree subdivision and it lends itself well to recursive implementation. You're not likely to get into stack-overflow problems with it, because with four times as many rects at each recursion cycle, you'll have created (potentially) over 1000 rects in just 5 cycles. Ten levels deep, you've created a million rects. Better start worrying about heap, not stack.

A demonstration of the technique can be seen below. The image on the left was created using a fairly "insensitive" variance threshold of 29 (meaning that subdivision did not occur unless a region had an RMS variance of at least 29, out of a possible range of pixel values of 0..255). Because subdivision happened infrequently, only in the very noisiest of pixel regions, smaller rects tended to cluster in areas of high detail (high variation in pixel intensities), such as around the edges of the iris and eyelashes. The image on the right shows that with the RMS threshold set lower (at 22), we get more detail -- about 300 rects total. (White outlines around the rects have been omitted on the right image.)


The image on the left has been parsed into 136 solid rectangles (with white outlines for illustrative purposes) using the recursive quadtree decomposition described in the text. Notice how subdivision into smaller rectangles tends to coincide with areas of high detail. The image on the right has been parsed into 334 rects.

The algorithm is tunable in a couple of ways. The most obvious way is via the variance-cutoff parameter. If the variance threshold is set low, it means that the slightest bit of "noise" in a given region will trigger subdivision of the region (and continued operation of the algorithm). However, we have to stop subdividing before reaching individual pixels. So another tuning variable is the minimum tile size.



The image on the left is a collage of 790 solid-filled rectangles. At a rect count of 2209, the image on the right is starting to have a bitmap-like feel.

The code that produced these images consists of 200 lines of JavaScript (shown below) and another 200 lines of Java utility routines (in a class called ImageUtilities: see further below). In addition to that, you need the ImageMunger Java application that I described in yesterday's post (yet another 200 lines or so of Java). First, the JavaScript:

// tiles.js
// Kas Thomas
// 23 January 2010
//
// Public domain.
//
// Note: For this to work, you need the ImageMunger
// Java app at http://3.ly/UBp
// You also need the ImageUtilities class described
// at the same blog.

// ------------------ RECURSIVE SUBDIVISION -------------------
function quadRecurse( ar, rect ) {

if ( !isDivisible( rect ) ) {
ar.push( rect );
return;
}

var newRects = quadDivide( rect ); // partition rect

for (var i = 0; i < newRects.length; i++) // size check
if (newRects[i][2] < 1 || newRects[i][3] < 1) {
ar.push(rect);
return;
}

for (var i = 0; i < newRects.length; i++) // recurse on each new rect
quadRecurse( ar, newRects[ i ] );
}

function quadDivide( rect ) {

var pixArray = getPixArrayFromRect( rect );

// Get the visual "center of gravity" of the image
var cg = Packages.ImageUtilities.getCG( pixArray, rect[2] );

cg[0] = (cg[0] + .5) * .5;
cg[1] = (cg[1] + .5) * .5;
cg[0] = 1.0 - cg[0];
cg[1] = 1.0 - cg[1] ;

var centerx = ( (cg[0] * rect[2]) & 0xffff);
centerx += rect[0];
var centery = ( (cg[1] * rect[3]) & 0xffff);
centery += rect[1];

var widthToCenterx = centerx - rect[0];
var heightToCentery = centery - rect[1];

var rect1 = [ rect[0], rect[1], widthToCenterx, heightToCentery ]; // UL
var rect2 = [ rect[0], centery, widthToCenterx, rect[3] - heightToCentery]; // LL
var rect3 = [ rect[0] + widthToCenterx, rect[1], rect[2] - widthToCenterx, heightToCentery ]; // UR
var rect4 = [ rect[0] + widthToCenterx, centery, rect[2] - widthToCenterx, rect[3] - heightToCentery ]; // LR

return [ rect1, rect2, rect3, rect4 ];
}

// -------------- divisibility ----------------
function isDivisible( rect ) {

if (rect[2] < WIDTH_THRESHOLD || rect[3] < HEIGHT_THRESHOLD)
return false;

var pixArray = getPixArrayFromRect( rect );
var rms = Packages.ImageUtilities.getRMSError( pixArray );

if (rms < RMSERROR_THRESHOLD)
return false;

return true;
}

function getPixArrayFromRect( rect ) {

var sub = Image.getSubimage( rect[0],rect[1],rect[2],rect[3] );
return sub.getRGB(0, 0, rect[2], rect[3], null, 0, rect[2]);
}

// -------------------- RENDER ---------------------------------------
function render( ar ) {

var g2d = Image.createGraphics();
var r = null; var sub = null; var pixels = null;
var color = null;

for (var i = 0; i < ar.length; i++) {
r = ar[i];
if (r[2] <= 0) continue; // r[2] = 1;
if (r[3] <= 0) continue; //r[3] = 1;

pixels = getPixArrayFromRect(r);
color = Packages.ImageUtilities.getAverageAWTColor( pixels );

g2d.setPaint( color );
g2d.fillRect( r[0], r[1], r[2], r[3] ); // FILL SOLID

if (OUTLINES == true) {
g2d.setColor( java.awt.Color.WHITE );
g2d.drawRect( r[0], r[1], r[2], r[3] );
}
}

Panel.updatePanel();
}

// -------------------- WRITE SVG ----------------------
function writeSVG( preamble, destfile, ar ) {
var r = null; var pixels = null; var awt = null;
var color = null;

var output =
'<?xml version="1.0" encoding="UTF-8" standalone="no"?>'+
'\n<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 20010904//EN" ' +
'"http://www.w3.org/TR/2001/REC-SVG-20010904/DTD/svg10.dtd">' +
'<svg xmlns="http://www.w3.org/2000/svg" ' +
'xmlns:xlink="http://www.w3.org/1999/xlink" ' +
'viewBox="0 0 640 480" ' +
'xml:space="preserve" ' +
'width="680px" ' +
'height="560px">';

output += "<g transform=\"scale(.85)\">";

for (var i = 0, r = null; i < ar.length; i++) {
r = ar[i];
pixels = getPixArrayFromRect(r);
awt = Packages.ImageUtilities.getAverageAWTColor( pixels );

color = awtColorToHex( awt );

output += outputSVGRect( mainArray[i], color );
}
output += "</g>";
output += "\n</svg>";

// write output to file
Packages.ImageUtilities.saveStringToFile(output, destfile);
}

function intToHex( num ) {
num *= 1;
hexStr = '000000' + (num).toString(16);
while ( hexStr.length > 6)
hexStr = hexStr.substring(1);

return "#" + hexStr;
}

function awtColorToHex( awt ) {
var theInt = (awt.getRed()<<16)|(awt.getGreen()<<8)|awt.getBlue();

return intToHex( theInt );
}

function outputSVGRect( r, color ) {

var str = "<rect x=";
str += "\"" + r[0] + "\" ";
str += "y=\"" + r[1] + "\" ";
str += "width=\"" + r[2] + "\" ";
str += "height=\"" + r[3] + "\" ";
str += "fill=\"" + color + "\" ";
str += "stroke=\"" + color + "\" ";
str += "/>\r";

return str;
}

// ---------- Main work routine ------------
// Usage:
//
// doQuadding( 10, 6, "svg", "C:/test1.svg" );
// (writes output to an SVG file)
//
// or:
//
// doQuadding( 11,4, "preview", null );
// (renders image in JFrame)

function doQuadding( rms, sizeLowLimit, mode, dest ) {

if (Image == null) {
java.lang.System.out.println("Nothing to do; no source image." );
return;
}
w = Image.getWidth(); h = Image.getHeight();
mainRect = [ 0,0,w,h ];
mainArray = new Array();

RMSERROR_THRESHOLD = rms;
WIDTH_THRESHOLD = HEIGHT_THRESHOLD = sizeLowLimit;

quadRecurse( mainArray, mainRect ); // *** RECURSE ***

java.lang.System.out.println("Total rects: " + mainArray.length );

if (mode.toLowerCase().indexOf("preview") != -1) {
java.lang.System.out.println(" rendering... " );
render( mainArray );
}
if (mode.toLowerCase().indexOf("svg") != -1) {
java.lang.System.out.println(" writing... " );
writeSVG( "c:/temp/svgStub.txt", dest, mainArray);
java.lang.System.out.println("DONE.");
}
}

OUTLINES = false;
var start = 1 * new Date;
// Actually call the entry point (begin processing):
doQuadding( 8,6, "preview", null );
// doQuadding( 8,6, "svg", "c:/temp/test86.svg" );
var end = 1 * new Date;
java.lang.System.out.println("Finished in " + (end-start) +
" milliseconds");

To use this file, give it a name like tiles.js, then run ImageMunger (the Java app I described in yesterday's post) from the command line, passing it the name of the image you want to modify, and the name of the script file:

> java ImageMunger myImage.jpg tiles.js

The entry point for the script is doQuadding(), which you can call with a 3rd argument of "svg" if you want to write output to a Scalable Vector Graphics file. Otherwise pass a 3rd arg of "preview" and ImageMunger will simply render the transformed image to a JFrame.

The script makes reference to a number of utility routines written in Java. The utility routines are in a class called (what else?) ImageUtilities, as follows. The routines should be fairly self-explanatory.

import java.io.BufferedReader;
import java.io.File;
import java.io.FileOutputStream;
import java.io.FileReader;
import java.io.IOException;
import java.io.OutputStream;

/*
ImageUtilities.java

Kas Thomas
23 January 2010
See http://3.ly/UBp and subsequent posts.
*/


public class ImageUtilities {

public static void saveStringToFile( String content,
String outpath ) {

OutputStream out = null;
try {
out = new FileOutputStream(outpath);
out.write(content.getBytes());
} catch (IOException e) {
System.out.println("Couldn't save to " + outpath);
e.printStackTrace();
} finally {
try {
if (out != null)
out.close();
} catch (IOException e1) {
e1.printStackTrace();
}
}
}

// Get the visual center-of-gravity of a pixel array.
// Pass the array and the raster width.
public static double [] getCG(int [] pix, int w) {

double intensity = 0;
int red = 0;
int green = 0;
int blue = 0;
double [] cg = { 0,0 };
double averageIntensity = 0;
int pvalue = 0;

for (int i = 0; i < pix.length; i++ ) {
pvalue = pix[i];
red = ((pvalue >> 16) & 255);
green = ((pvalue >> 8) & 255);
blue = pvalue & 255;
intensity = ((double)(red + blue + 2 * green))/1024.;
averageIntensity += intensity;
cg[0] += intensity * (i % w);
cg[1] += intensity * (i / w);
}

cg[0] /= averageIntensity;
cg[1] /= averageIntensity;

cg[0] /= w;
cg[1] /= pix.length/w;
return cg;
}

public static double getRMSError( int [] pix ) {

double accumulator = 0;
double diff = 0;
double aveIntensity = 0;
double rms = 0;
int len = pix.length;

for (int i = 0; i < len; i++) {
aveIntensity += (double)((pix[i] >> 8) & 255);
}

aveIntensity /= len;

for (int i = 0; i < len; i++) {
diff = (double)((pix[i] >> 8) & 255) - aveIntensity;
accumulator += diff * diff;
}

rms = accumulator/len;

return Math.sqrt(rms);
}

public static java.awt.Color getAverageAWTColor( int [] input ) {

int ave = getAverageColor( input );
int red = (ave >> 16) & 255;
int green = (ave >> 8) & 255;
int blue = ave & 255;
return new java.awt.Color(red,green,blue);
}

public static int getAverageColor( int [] input ) {

int red = 0;
int green = 0;
int blue = 0;
int pvalue = 0;
int averageRed = 0;
int averageGreen = 0;
int averageBlue = 0;

int len = input.length;

for (int i = 0; i < len; i++) {

pvalue = input[i];
red = ((pvalue >> 16) & 255);
green = ((pvalue >> 8) & 255);
blue = pvalue & 255;

averageRed += red;
averageGreen += green;
averageBlue += blue;

}

averageRed /= len;
averageGreen /= len;
averageBlue /= len;

return (averageRed << 16) | (averageGreen << 8) | averageBlue;
}

public static double getIntensity( int pvalue ) {

int red = ((pvalue >> 16) & 255);
int green = ((pvalue >> 8) & 255);
int blue = pvalue & 255;

double intensity = red + blue + 2 * green;

return intensity/1024.;
}

public static double getIntensity( java.awt.Color c) {

int intvalue = c.getRed() << 16;
intvalue += c.getGreen() << 8;
intvalue += c.getBlue();
return getIntensity( intvalue );
}

public static java.awt.Color getAWTColor( int pvalue ) {

int red = ((pvalue >> 16) & 255);
int green = ((pvalue >> 8) & 255);
int blue = pvalue & 255;

return new java.awt.Color(red,green,blue);
}

public static double getRMSE( int [] pix1, int [] pix2) {

double rms = 0;
double accum = 0;
double intensity1 = 0;
double intensity2 = 0;
double tmp = 0;

if (pix1.length != pix2.length) {

System.out.println("Arrays are not the same size.");
return rms;
}

for (int i = 0; i < pix1.length; i++) {
intensity1 = getIntensity( pix1[i] );
intensity2 = getIntensity( pix2[i] );
tmp = intensity1 - intensity2;
tmp *= tmp;
accum += tmp;
}

rms = accum/pix1.length; // the mean of the squares
return Math.sqrt( rms ); // root mean square
}
}

Even though the main routine is in JavaScript, the overall processing runs quickly. The algorithm executes in near-linear time and can output around 5000 rectangles per second (to screen, that is; disk I/O not included).

Projects for the future:
  • Write gradient-filled rects instead of solid-filled rects, choosing the gradient endpoint colors in such a way as to minimize the differences between the output and the original image.
  • Given two SVG images, each an ensemble of rects (preferably equal in number), write a transformation routine that morphs one image into the other via transformations to the individual rects (and their colors). Store the second image as simply an ensemble of transformations (no rects). The first image provides the "reference rectangles" that will be transformed.
  • Write a public key image-encryption routine based on the foregoing notion, such that Image A becomes a key that someone can use to encrypt Image B, with private-key Image C unlocking B.
  • Instead of writing an SVG image as an ensemble of individual rects, write it as one rect that is repeatedly resized and repositioned via successive affine transformations.
  • Encrypt an image using the above technique (rewrite one rect many times through transformation) in concert with matrix hashing. Write an SVG image-encryption routine whose difficulty of decryption depends on the non-commutative nature of matrix multiplication and the numerical instability of inverted matrices.
  • Write a "chunky JPEG" routine that essentially uses the quadtree decomposition to chunk the image prior to discrete cosine transformation, instead of using the (canonical JPEG) 8x8 chunking.
  • [ Insert your own idea here! ]
reade more... Résuméabuiyad

A simple Java class for running scripts against an image

I like to do a bit of 2D and 3D graphics programming in my spare time. The trouble is, Java gets a bit heavy at times, and I look for ways to test new ideas quickly. Testing ideas quickly means scripting. But how do you get at pixel values with JavaScript?

The answer is, you write a super-simple Java class that can load an image and hand off pixels to the JavaScript context. The following is such a class.
/*
ImageMunger
Kas Thomas
http://asserttrue.blogspot.com/

Public domain code.

All it does is open an image, display it in a window,
and run a script file. Usage:

ImageMunger myimage.jpg myscript.js

The JavaScript file will (at runtime) have two global variables,
Panel and Image, in scope. They correspond to the
ImagePanel inner class instance and the BufferedImage
that's displaying in the panel. That way, your script
can easily manipulate the image and/or the panel.

*/

import java.awt.Graphics;
import java.awt.image.BufferedImage;
import java.io.BufferedReader;
import java.io.File;
import java.io.FileOutputStream;
import java.io.FileReader;
import java.io.IOException;
import java.io.OutputStream;
import javax.imageio.ImageIO;
import javax.swing.JComponent;
import javax.swing.JFrame;
import javax.script.ScriptEngine;
import javax.script.ScriptEngineManager;

public class ImageMunger {

// This inner class is our canvas. We draw the image on it.
class ImagePanel extends JComponent {

BufferedImage theImage = null;

ImagePanel( BufferedImage image ) {
super();
theImage = image;
}

public BufferedImage getImage( ) {
return theImage;
}

public void setImage( BufferedImage image) {
theImage = image;
this.updatePanel();
}

public void updatePanel() {

invalidate();
getParent().doLayout();
repaint();
}

public void paintComponent( Graphics g ) {

int w = theImage.getWidth( );
int h = theImage.getHeight( );
g.drawImage( theImage, 0,0, w,h, this );
}
} // end ImagePanel inner class

// We need to keep a reference to the ImagePanel:
public ImagePanel theImagePanel = null;

// Constructor
public ImageMunger( String [] args ) {

parseArgs( args );

// open image
BufferedImage image = openImageFile( args[0] );

// create window
theImagePanel = new ImagePanel( image );
JFrame gMainFrame = new JFrame();
gMainFrame.setTitle( args[0] );
gMainFrame.setBounds(50,80,
image.getWidth( )+10, image.getHeight( )+10);
gMainFrame.setDefaultCloseOperation(3); // dispose
gMainFrame.getContentPane().add( theImagePanel );
gMainFrame.setVisible(true);

}

ImagePanel getImagePanel( ) {

return theImagePanel;
}

BufferedImage openImageFile( String fname ) {

BufferedImage img = null;

try {
File f = new File( fname );
img = ImageIO.read(f);
}
catch (Exception e) {
showMessage("Trouble reading file.");
e.printStackTrace();
}

return img;
}

public static void runScriptFromFile( String fileName,
ScriptEngine engine ) {

try {
engine.eval(new java.io.FileReader( fileName ));
}
catch( Exception exception ) {
exception.printStackTrace();
showMessage( exception.getMessage() );
}
}


public static void showMessage(String s) {
javax.swing.JOptionPane.showMessageDialog(null, s);
}

void parseArgs( String[] args ) {

if ( args.length < 2 )
tellUserHowToUseThisApp( );
}

void tellUserHowToUseThisApp( ) {
showMessage( "Supply an image file name and a script file name." );
}

// main()
public static void main( String[] args ) {

ImageMunger munger = new ImageMunger( args );

// create a script engine manager & engine
ScriptEngineManager factory = new ScriptEngineManager();
ScriptEngine engine = factory.getEngineByName("JavaScript");

engine.put( "Image", munger.getImagePanel( ).getImage( ) );
engine.put( "Panel", munger.getImagePanel( ) );
engine.put( "args" , args );

// evaluate JavaScript code from file
runScriptFromFile( args[1], engine );
}
}

This Java app does two things, based on arguments provided on the command line. First, it opens an image of your choice (you supply the file path as the first command-line arg) in a JFrame. Secondly, it opens and executes the JavaScript file whose path you supply as the second command-line argument.

Before executing the script, the Java app puts two variables, Image and Panel, into the JavaScript runtime scope. (These are references to the BufferedImage object, and the JComponent that renders it, respectively.) That way, it's easy to manipulate the image and/or its JComponent at runtime, using script.

Here's an example. Suppose you want to convert a color JPEG into a grayscale (black and white) image. You would write a script like the following and put it in a .js file on disk. Then run the ImageMunger app and have it execute the script.


// desaturate.js
// Converts RGB image to grayscale
.

( function desaturate( ) {

if (Image == null) {
java.lang.System.out.println("Nothing to do; no source image." );
return;
}

var w = Image.getWidth();
var h = Image.getHeight();
var pixels = Image.getRGB( 0,0,w,h,null,0,w );
var tmp;

for ( var i = 0; i < pixels.length; i++ ) {

// get Green value
tmp = 255 & (pixels[ i ] >> 8);

// set Red, Green, and Blue channels the same
pixels[ i ] = (tmp << 16) | (tmp << 8) | tmp;
}

Image.setRGB( 0,0,w,h,pixels,0,w );
Panel.updatePanel( );
} ) ( );

No real magic here. The script just grabs pixels from the BufferedImage (via the Image global that we put into the JS scope), converts all pixels in the image to monochrome RGB values, stuffs the pixels back into the image, and requests a redraw of the ImagePanel (using the Panel global).

No rocket science. Just quick-and-dirty pixel banging. Perfect if you're into peeking and poking pixels late at night but like to avoid getting Java under your fingernails just before climbing into a clean bed.
reade more... Résuméabuiyad