distance transforms

dt = bwdist(img [,algorithm])

- img
- Binary image. If every pixel is different than 0, the output is undefined and may contain arbitrary values.
- algorithm
- Listed below are various algorithms that are available in SIP,
together with the shortest string form accepted (for speed
of use). This argument is CASE-INSENSITIVE, for convenience. Some of the
algorithms are faster than others, but this depends heavily
on the size and content of the input.
- 'euclidean' or 'euclid'
- the default fast exact euclidean algorithm. Currently, it is the same as the 'maurer' bellow. (DEFAULT)
- 'maurer' or 'mau'
- very fast (and recent) exact euclidean algorithm, based on certain dimensionality properties of Voronoi diagrams. It is the best method together with Meijster, in general.
- 'meijster' or 'mau'
- very fast (and recent) exact euclidean algorithm, based on certain dimensionality properties of Voronoi diagrams. It is the best method together with Maurer, in general. For most cases it is slightly faster than Maurer, but uses a little more memory.
- 'cuisenaire pmn'
- very fast exact euclidean algorithm. It is based on propagation of multiple neighborhoods to build up an exact EDT.
- 'cuisenaire pmon'
- a variation of the latter that uses multiple oriented neighborhoods It seems to be slightly slower, in general, but can be faster for some cases (we don't know exactly which)
- 'cuisenaire psn4'
- a variation of the latter that uses only 4-neighborhood. This is faster but less precise
- 'cuisenaire psn8'
- a variation of the latter that uses diagonal neighborhood after 4-neighborhood to improve the precision. This is faster than the full 'pmn' algorithm, but less precise. It is a little slower than psn4 but considerably more precise.
- 'lotufo-zampirolli' or 'lotufo-z'
- very fast exact euclidean algorithm. Seems to be slightly slower than maurer and cuisenaire, in general, but can be faster for some cases.
- 'IFT 8' or 'IFT'
- a fast algorithm using the euclidean metric. For large and thick shapes, there may be a few small errors, which are dispensable for most practical applications.
- 'IFT 4'
- the same algorithm but with 4-neighborhood propagation. This means that this method is about 2x faster but less precise
- 'exact dilations' or 'exact dil'
- will perform an exact euclidean algorithm that is slow for medium shapes, but it is always exact and reasonably fast for thin shapes.

- dt
- The distance transform of the image. It is undefined if the input is an image without any pixels equal to 0.

Function bwdist computes the distance transform. For each foreground pixel (i.e. value '1') in the input image, the distance transform assigns a value that is the smallest distance between that pixel and the all the 0-pixels (you can also think of the distance to the outer boundary of the object).

Many different methods are provided for comparison purposes. If you are going to use bwdist extensively, you may test all the algorithms to find the best one for your particular type of image.

xset('auto clear', 'on'); // First, a simple example to illustrate the concept A = zeros(15,10); A(4:12,3:7)=1; A(4:5,3:4)=0 D = bwdist(A) D = sqrt(D) // Note how the values in D were calculated. // For each pixel p such that A(p)=1, D(p) is the minimum euclidean // distance of p and the 0-pixels (background). // ----------------------------------- // Now to a more interesting example // ----------------------------------- A = gray_imread(SIPDIR + 'images/escher.png'); imshow(A,2); D = bwdist(A); imshow(log(1+D),[]); // normalizes image to enhance visualization D = bwdist(A,'exact dilations'); imshow(log(1+D),[]); // To obtain an external EDT, simply invert the shape: B = 1-A; D = bwdist(B,'maurer'); imshow(log(1+D),[]); // To obtain an external+internal EDT, simply compute // the binary border of the shape and pass its negative // to bwdist: A = bwborder(A); A = 1-A; D = bwdist(A,'lotufo-zampirolli'); imshow(log(1+D),[]); // --------------------------------- // Other forms to visualize the DT // --------------------------------- // Wrapping (note the wavefronts of iso-distance) imshow(modulo(sqrt(D),10),[]) // Usual: D = bwdist(A); D = normal(sqrt(D),1000,1); imshow(D,hotcolormap(1000)); // There is also of DT application in the example for the "watershed" // function. xset('auto clear', 'off'); |

**The original image and its distance map:**

**The distance map of the border (yielding
non-zero distances inside and outside), and
the propagating wavefronts:**

- Maurer
Maurer, C.R. and R. Qi and V. Raghavan, "A Linear Time Algorithm for Computing the Euclidean Distance Transform in Arbitrary Dimensions", IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no. 2, pp. 265-270, february 2003.

- Meijster
A. Meijster, J.B.T.M. Roerdink, and W.H. Hesselink "A General Algorithm for Computing Distance Transforms in Linear Time", proc. of 5th Int. Conf. Mathematical Morphology and its Applications to Image and Signal Processing, 2000

- Cuisenaire
Cuisenaire, O and Macq, B, "Fast Euclidean Distance Transformation by Propagation Using Multiple Neighborhoods", Computer Vision and Image Understanding, no. 2, vol 76, 163--172, 76, 1999.

Chapter 3 of "Distance transformations: fast algorithms and applications to medical image processing", Olivier Cuisenaire's Ph.D. Thesis, October 1999, Université catholique de Louvain, Belgium.

- IFT
"Multiscale Skeletons by Image Foresting Transform and its Application to Neuromorphometry", A.X. Falcao, L. da F. Costa, B.S. da Cunha, Pattern Recognition, 2002.

- Lotufo-Zampirolli
R. Lotufo and F. Zampirolli, Fastmultidimensional parallel euclidean distance transform based on mathematical morphology, in T. Wu and D. Borges, editors, Proccedings of SIBGRAPI 2001, XIV Brazilian Symposium on Computer Graphics and Image Processing, pages 100-105. IEEE Computer Society, 2001.

- Exact Dilations
"Multiresolution shape representation without border shifting", L. da F. Costa, and L. F. Estrozi, Electronics Letters, no. 21, vol. 35, pp. 1829-1830, 1999.

"Shape Analysis and Classification", L. da F. Costa and R.M. Cesar Jr., CRC Press.

- Ricardo Fabbri <ricardofabbri (AT) users DOT sf DOT net>

http://siptoolbox.sourceforge.net