WATERSHED TRANSFORMATION

1
WATERSHED TRANSFORMATION

• M.SACHIDANAND
• October 15th,2007

2
Image Segmentation

• Process of partitioning a digital image into
  multiple regions (sets of pixels)
• Simplify and/or change the representation of an
  image into something that is more meaningful and
  easier to analyze.
• Typically used to locate objects and boundaries
  (lines, curves, etc.) in images.

3
WATERSHED TRANSFORMATION

• Image data may be interpreted as a topographic
  surface where the gradient image gray-levels
  represent altitudes.
• Region edges correspond to high watersheds and
  low-gradient region interiors correspond to
  catchment basins.

4
WATERSHED TRANSFORMATION

• Catchment basins of the topographic surface are
  homogeneous in the sense that all pixels
  belonging to the same catchment basin are
  connected with the basin's region of minimum
  altitude (gray-level) by a simple path of pixels
  that have monotonically decreasing altitude
  (gray-level) along the path.
• Such catchment basins then represent the regions
  of the segmented image.

Original Image
Watershed Segmented Image
5
Watershed Based Image Segmentation Procedure
6
APPLICATIONS

• Medical Imaging
• Semiconductor Technology
• Traffic Monitoring
• Face Recognition
• Identifying Fractures in Steel

Coffee Beans Separation
7
Pattern Recognition Letters 26 (2005) 12661274

• A fast watershed algorithm based on chain code
  and its application in image segmentation
• -Han Sun , Jingyu Yang, Mingwu Ren

8
ABSTRACT

• A new Watershed Transformation Algorithm based on
  chain code.
• Complexity of the proposed algorithm is
  discussed.
• Experiment results are shown to prove that the
  presented algorithm runs fast.

9
CHAIN CODES

• A chain code is a more concise way of
  representing a list of points than a simple
  (x1,y1),(x2,y2) ..... .
• A chain code describes the boundary as a series
  of one pixel vector transitions at different
  orientations only the starting point is defined
  explicitly.
• Chain code is thought as representing the contour
  of the image.
• Each code can be considered as the angular
  direction, in multiples of 45deg., that we must
  move to go from one contour pixel to the next.

10
CHAIN CODES
11
Definitions for chain code basedWatershed
Transformation

• Point-out chain code is the directional code that
  current pixel points to.
• 2) Point-in chain code is the directional code
  that neighbor pixel points to.

12
Steps involved in the Algorithm

• Step-A For each pixel in a given image, its
  lowest neighborhood is detected.

13
Steps involved in the Algorithm

• 2) Step-B For each pixel on plateaus, if one
  of its neighbors points somewhere (lower
  plateau), the pixel points the neighbor.
• Repeating this procedure using a FIFO, pixels on
  plateaus point one of their lower plateaus. The
  FIFO is used to divide plateaus evenly according
  to the distances from their lower plateaus.

14
Steps involved in the Algorithm

• 3) Step-C For each pixel which points, a unique
  label is given. Then, its neighbor pixels on the
  same plateau point the pixel. By repeating this
  procedure, all pixels on the plateau points the
  pixel directly or indirectly.
• 4) Step-D Pointers for all pixels are
  dereferenced, and labels are given to all pixels.

15
EXPERIMENTS AND RESULTS

• Watershed transformation and region merging
  (brain)
• (a) brain (388 395)
• (b) gradient image, m 3
• (c) watershed results(6227 regions)
• (d) merging results (11 regions).

16
EXPERIMENTS AND RESULTS

• Watershed transformation and region merging
  (blood)
• (a) blood (272 265)
• (b) gradient image, m 3
• (c) watershed results(1644 regions)
• (d) merging results (27 regions).

17
Performance Analysis on Pentium IV 2.4 GHz CPU

• The above 3 algorithms produce the same number of
  regions, hence identical in function.
• Algorithm BM and new(chain code) algorithms are
  faster.

18
IMPLEMENTATION OF A PARALLEL AND PIPELINED
WATERSHED ALGORITHM ON FPGA -Dang Ba Khac
Trieu and Tsutomu Maruyama
International Conference on Field Programmable
Logic and Applications, 2006
19
ABSTRACT

• Implementation of a parallel and pipelined
  watershed algorithm on FPGA.
• Implemented on XC2V6000 and up to 32 pixels are
  processed in parallel.
• The performance for 512 x 512 pixel images is
  found to be about 3 - 4 msec, which is fast
  enough for real-time applications.

20
4 Basic Steps in this process

• A given image is divided to K regions.
• Several of them are cached on the FPGA.
• The watershed algorithm is applied on those
  regions.
• 4) The next (or previous) region is loaded to the
  FPGA during the computation to hide the loading
  time.

21
PIPELINED AND PARALLEL ALGORITHM

• 1) Two arrays are used.
• 2) One array to store labels. Another to divide
  non-minima plateaus evenly.
• 3) Repetitive sequential scanning of pixels from
  top-left to bottom-right and vice-versa.
• 4) Repetition is finished when no values of array
  are modified in the current scanning.

22
Pseudocode
23
Steps-1 and 2 in Algorithm
24
Steps-1 and 2 in Algorithm
25
Step-3 in Algorithm
26
Step-3 in Algorithm
27
Step-3 in Algorithm
28
IMPLEMENTATION OF THE ALGORITHM

• 1) Without Cache Memory on FPGA
• All data are stored in external Memory Banks.
• Performance varies with the number of memory
  banks.

29
IMPLEMENTATION OF THE ALGORITHM-Parallel and
Pipelined Processing

• Four lines are processed in parallel.
• We need to wait 3 clock cycles before starting
  the computation of the next line.
• The total clock cycles to processes N lines
  becomes
• L (N - 1) 3 where L is the pipeline
  depth.

30
IMPLEMENTATION OF THE ALGORITHM

• 2) With Cache Memory on FPGA
• A given image is divided to smaller regions along
  x (or y) axis.
• Algorithm is applied to a region cached in some
  of these cache memory banks.
• Data for next region is downloaded to another
  cache during this computation

31
IMPLEMENTATION OF THE ALGORITHM Trace of
Computation with Cache on FPGA
32
EXPERIMENTAL RESULTS

• The algorithm was implemented on ADM-XRC-II
  with one XC2V6000.
• X or Y of the given image can not be larger
  than 512.
• Performance of circuit is faster than 30 frame
  per second with one unit.
• With four units (without cache memory banks),
  the performance
• can be improved about 3.7 - 3.8 times.

33
EXPERIMENTAL RESULTS
34
LIMITATIONS

• Performance gain by 32 units over four units is
  limited.
• Reasons
• 1) In the approach with cache memory banks, the
  algorithm is applied at least twice to each
  region.
• 2) Number of back traces is huge.
• 3) Pipeline overhead is high because of usage of
  32 parallel lines.

35
FUTURE WORK

• To improve performance gain attained by 32 units
  over four units.
• - reducing pipeline overhead.
• - using column optimized scanning.