Depending on the width of the image a lot of int[] have been created, a BitSet with a size of width * height is a better choice here. It uses less memory and is faster than determinating a pixels transparency each time.

Create a copy of the original raster to do the shouldPad test and instant update the raster. This avoids a lot of object creation, hash calculation and garbage collection needed to maintain the HashSet.

It is more efficent to check the boudings outside the loop

should be a little faster

we expect the none transparent pixel to be around the center.
we check a 5 * 5 raster first to move faster to the center.
after this we still check all pixels to get a complete result.

if padding is used there are a lot of transparent lines

…s raster

this improves the image quality and removes intermediate step

sorry, commited a tmp image

the concurrent creation of words could break client code

the graphics object used to create the font metrics had other RenderingHints than the graphics object to write the text, if the default values of the used jvm did not have the TEXT_ANTIALIASING set to TEXT_ANTIALIAS_LCD_HRGB.

…o it

the maximum spiral radius might be bigger or smaller than the image width, e.g. if the image height is bigger than the width.

this happened when the offset was 0