Image Processing - Segmentaion

object detection

• no exact outlining

• useful for counting

semantic segmentation

• find region for obect type

• useful for calculating coverage

instance segmentation

• find individual objects of different types

• counting and coverage

standard approach of foreground/background segmentation

very low computational cost

determine for each pixel based on value (threshold)

local threshold based on mean within window

approach to seperate merged objects

fill up image with water from the “bottom” (lowest intensity)

create barrier at position, where water of different basins meet

Connect/disconnect objects without overall distortion that pure dilation/erosion gives

Opening -> erosion followed by dilation

Closing -> vice versa

form graph with one pixel for every node and add source and sink node

cut graph, based on max flow

edges represent cost for both images being in one class

identify object in given scene -> works best if template is in scene

assign matching error to each pixel in image

in biology, approximately circular objects need to be found often

similar to template matching, convolve image with blob-like structure

f.e. Laplacian of Gaussian, derivation of Gaussian

allign active contour along boundary of object, based on minimizing energy

active contour energy divided into internal energy and external energies

internal energy: elasticity and stiffness of contour

collection of pixel, that have similar properties (color, intensity) and are spatially close

superpixel segmentation intermediate step

label number of pixels in image

based on random walk, for every unlabeled pixel, compute which is the first labeled pixel that reaches it

assign label of this pixel to the unlabeled pixel